About
Activity
2K followers
-
Pradeep Jindal shared thisThe AI components of a voice agent are the easy part. Here's what actually breaks. Telephony lies to you. SIP error codes don't mean what the RFC says they mean. We tripled our error mappings to nearly 20 based purely on observed production behavior. Audio quality on real phone networks breaks your STT pipeline - carrier codec compression, spotty connectivity, and real-world background noise produce transcription variants you'd never invent in testing. Silence is ambiguous - users think, get distracted, put the phone down, and your agent needs to handle all of these differently. Interruption handling turns out to be a cultural problem: in Hindi conversation, "haan" and "achha" are listening sounds, not interruptions, and duration-based detection cuts the agent off on every one. None of these show up in demos. All of them determine whether your deployment succeeds. Here's our final post of the series on voice agents, production hardening: https://lnkd.in/d2pf9Xy8 #VoiceAI #ProductionEngineeringThe Parts of Voice Agents That Break in ProductionThe Parts of Voice Agents That Break in Production
-
Pradeep Jindal shared thisWe built cost tracking to bill clients. It became the best bug detector we had. Mid-campaign, per-call cost jumped from ~$0.018 to ~$0.045 while average call length barely moved - the same conversations, suddenly burning 2.5x the tokens. Every health check was green and no dashboard was red; the cost line was the only thing that flagged it. The agent had hit a phrase it couldn't parse and was looping, regenerating longer responses every turn without making progress. A billing metric caught a behavioral regression before anyone reviewing transcripts did. This is why voice observability has almost nothing to do with traditional APM. Your service returns 200 OK while the agent hangs up on a frustrated customer because it read a complaint as a goodbye. Uptime tells you the system is running; it tells you nothing about whether the conversation worked - and in voice you can't retry, so a call you didn't catch going wrong is just gone. You have to see inside the call: per-turn latency from VAD to TTS, full transcripts, post-call analysis, and cost per call as a leading indicator. So we built three layers for three audiences - per-turn metrics for engineers, lifecycle tracking for operations, post-call AI analysis for the business. Here's how the whole system fits together. https://lnkd.in/gTsS5NsD #VoiceAI #ObservabilityVoice Agent Analytics & Observability: From Raw Calls to Business IntelligenceVoice Agent Analytics & Observability: From Raw Calls to Business Intelligence
-
Pradeep Jindal shared thisA TTS engine that sounds natural on general text but reads a product code as an ordinal number is unusable. An STT provider with great benchmark WER that transcribes your most common brand name six different ways is unusable. A pipeline change that shaves 200ms off STT but introduces enough transcription errors to create correction turns that add more latency than it saved - also unusable. We learned the hard way that testing what the agent says is only one layer. We now test five independently: LLM response correctness, STT provider accuracy on real telephony audio, TTS pronunciation for names and numbers, RAG retrieval accuracy with boundary detection, and end-to-end latency benchmarking where every pipeline change runs through a standardized harness before deploying. Our most valuable response tests still come from production bugs turned into permanent regression cases - after two campaigns, those caught more issues than our hand-written suite. We wrote up how we test each layer and what we still haven't solved. https://lnkd.in/dj3Z-8yw #VoiceAI #AITesting
-
Pradeep Jindal shared thisOur production voice agent system prompts total 2,700+ lines. Every line earned its place through a real call failure. A 200-line prompt works in demos. Then a user mentions a product and the agent helpfully adds the brand name they never said. A user expresses dissatisfaction and the agent ends the call, interpreting negativity as goodbye. A user speaks a numeric code using Hindi conventions and the agent hears it as an ordinal. Voice prompt engineering is fundamentally different from chat. Hallucination is audibly destructive - the user hears it immediately and loses trust. Negative feedback looks like goodbye to an LLM, but it's the most valuable data you're collecting. Domain vocabulary is infrastructure - a single brand name can have six phonetic spellings in Hindi STT output. We chose prompts over fine-tuning deliberately: deploy in minutes, express precise rules, audit every line. We wrote up the patterns that got us from 200 lines to 2,700. Next one in our series: https://lnkd.in/gGJz9hTt #VoiceAI #PromptEngineering
-
Pradeep Jindal shared thisOur most expensive voice agent call costs $0.046. A human agent doing the same work: $0.50-2.00. Real per-call numbers are nearly impossible to find online, so we published ours from thousands of production calls. The most counterintuitive lesson: our STT provider that costs 25% more per minute is actually cheaper per call. Better Hindi accuracy means fewer misunderstandings, fewer clarification loops, shorter calls. We've never needed a flagship LLM for any voice agent task - mini-tier models handle structured conversations at 10-17x lower cost. Prompt caching cuts input token costs ~50% with zero engineering effort. But the single biggest cost lever isn't any AI component. It's call quality. A call that collects good data in 90 seconds is cheaper than one that takes 4 minutes because the agent misunderstood twice. At two cents per call, shaving another 10% is negligible. Invest in the conversation instead. Read about the per-component breakdown, what we optimized, and what we deliberately chose not to. Next one in our series: https://lnkd.in/gFX8x_nx #VoiceAI #AIAgentsVoice Agent Cost Optimization: Real Numbers from ProductionVoice Agent Cost Optimization: Real Numbers from Production
-
Pradeep Jindal shared thisWe spent three months tuning latency on a single voice agent deployment. And reversed our own decisions multiple times. Disabled ML-based turn detection because of high latency. Conversations got choppy. Re-enabled it with an aggressively tuned threshold. That worked. Tried preemptive LLM generation to overlap inference with user speech. Backfired for multilingual conversations - partial transcriptions in code-switching languages are unreliable. Disabled. There is no magic configuration. Every parameter interacts with every other parameter. The only path is systematic measurement and iteration. Which is why per-turn instrumentation - VAD, STT, LLM TTFT, TTS TTFB for every single conversational turn - isn't a nice-to-have. It's the prerequisite for all optimization. We wrote up the full playbook - every decision we made, reversed, and why. This is the third blog post in our series. https://lnkd.in/g8cjEWkr #VoiceAI #AIAgentsVoice Agent Latency: The Sub-Second Tuning PlaybookVoice Agent Latency: The Sub-Second Tuning Playbook
-
Pradeep Jindal shared thisWe evaluated voice AI platforms before our first production deployment. They get you to a demo in hours. Then you discover the problems that actually determine whether a deployment succeeds. We needed a silence detection plugin with an 8-second threshold (started at 3-4 seconds - far too aggressive), escalating nudge prompts in Hindi, grace periods that account for agent processing time, and a nudge counter that resets only on user speech. We needed to swap our STT provider mid-deployment because a single brand name was getting transcribed six different ways. The specialized Indian-language provider costs 25% more per minute but better accuracy means fewer misunderstandings, shorter calls, and lower cost per call. The real question isn't "platform or custom." It's where your production problems live relative to the abstraction boundary. We ended up with per-call costs 2-10x lower than platforms, but the cost savings aren't why we built it. We built it because the solutions we needed required control that platforms don't expose. We wrote up the full decision framework, with real cost comparisons. https://lnkd.in/gEZWsnpd #VoiceAI #AIAgentsBuild vs. Buy for Voice AI: Why We Chose Neither (and Both)Build vs. Buy for Voice AI: Why We Chose Neither (and Both)
-
Pradeep Jindal posted thisAfter 20+ years in tech, I feel like I've just started again. Curious. Slightly uncomfortable. Relearning. And honestly - that feels right. Here's why. Most of us knew the rat race was exhausting. We just didn't stop running. The system rewarded skill-stacking. Certifications. Knowing more than the room. Being the "smartest engineer" in the meeting. So we kept going. Quietly competent. Quietly defensive. Quietly hoping something would force a change we couldn't make ourselves. AI did. Skills are becoming cheaper. Information is abundant. Execution is accelerating. The edge many of us spent decades building isn't gone - but it's no longer sufficient on its own. You can't fake this stuff anymore: whether you can say "I might be wrong" and mean it. Whether you're building for the work or for the image. AI didn't just automate tasks. It called the bluff on expertise. After Yahoo, InMobi, a few startups, and now building ByondLabs - I feel like I've just hit reset. Excited and nervous at the same time, if I'm honest. The race worth running is the one that starts inside - clarity of thinking, intellectual honesty, building for truth over ego. The old race had no finish line. Neither does this one. The difference is - this one is worth running. (Credit: My 14-year-old daughter read a draft of this and restructured it better than I had.)
-
Pradeep Jindal shared thisBuilding a voice AI demo takes an afternoon. Getting it to handle real phone calls - across regional accents, mixed languages, noisy environments, and the full spectrum of telephony edge cases - took us months. We've shipped production voice agents handling thousands of calls in Hindi, Hinglish, and English. Average cost: ~$0.03 per call. The hard part wasn't the AI. It was everything around it - turn detection that works for Hinglish speakers, silence thresholds calibrated from hundreds of real calls, SIP error mappings no RFC prepared us for, and 2,700+ lines of prompt engineering where every section traces to a real call failure. We wrote up what we learned. This is the first post in a series covering the full landscape: architecture, latency, cost, prompt engineering, testing, production hardening, and observability. If you're building or evaluating voice AI, this might save you some months. https://lnkd.in/gGZuU6be #VoiceAI #ConversationalAI #AIAgentsWhat It Takes to Build Production-Grade Voice AI AgentsWhat It Takes to Build Production-Grade Voice AI Agents
-
Pradeep Jindal liked thisPradeep Jindal liked thisThe Architecture of Alpha: From Neural Inference to Live Execution In the current era of Generative AI, writing code is no longer the bottleneck. Syntax is free; the barrier to entry has collapsed. The only sustainable edge left is architecture and domain-specific logic. Over the last year, I built the end-to-end machine learning and quantitative execution engine for retailtrader.ai. I wanted to solve a notoriously difficult problem: identifying high-probability structural chart patterns before they become obvious, and executing them without look-ahead bias or latency penalties. The era of "I trained an LSTM in a Jupyter notebook" is over. Models fail in production because the underlying pipeline suffers from data drift, orchestration bottlenecks, and live market slippage. I am launching a 5-part series breaking down the complete engineering infrastructure behind this predictive engine—from the neural network backpropagation down to the Apache Hudi data lakehouse and the MySQL execution state machine. Here is the roadmap for the upcoming series: Part 1: The Cloud Blueprint & Serverless Data Lakehouse. How we separated high-throughput time-series analytics from low-latency state management using AWS Fargate, Step Functions, Apache Hudi, and Serverless Athena. Part 2: The Neural Backbone. Why we discarded lagging technical indicators for pure price microstructure using a Dilated CNN-LSTM, Focal Loss (to handle extreme class imbalance), and Federated Averaging (FedAvg). Part 3: Signal Processing & Zero Look-Ahead Bias. Turning continuous probabilities into discrete trade setups using probability momentum and in-memory edge padding. Part 4: Kinematic Strategy Logic. Using Athena to engineer derived features like SMA trend velocity/acceleration, and enforcing hard-coded structural regimes (Golden Crosses, dynamic Risk/Reward). Part 5: The Execution Engine & Portfolio State Machine.retailtrader.ai. I wanted to solve a notoriously difficult problem: identifying high-probability structural chart patterns before they become obvious, and executing them without look-ahead bias or latency penalties. The era of "I trained an LSTM in a Jupyter notebook" is over. Models fail in production because the underlying pipeline suffers from data drift, orchestration bottlenecks, and live market slippage. I am launching a 5-part series breaking down the complete engineering infrastructure behind this predictive engine—from the neural network backpropagation down to the Apache Hudi data lakehouse and the MySQL execution state machine. Here is the roadmap for the upcoming series: Part 1: The Cloud Blueprint & Serverless Data Lakehouse. How we separated high-throughput time-series analytics from low-latency state management using AWS Fargate, Step Functions, Apache Hudi, and Serverless Athena. Part 2: The Neural Backbone. Why we discarded lagging technical indicators for pure price microstructure using a Dilated CNN-LSTM, Focal Loss (to handle extreme class imbalance), and Federated Averaging (FedAvg). Part 3: Signal Processing & Zero Look-Ahead Bias. Turning continuous probabilities into discrete trade setups using probability momentum and in-memory edge padding. Part 4: Kinematic Strategy Logic. Using Athena to engineer derived features like SMA trend velocity/acceleration, and enforcing hard-coded structural regimes (Golden Crosses, dynamic Risk/Reward). Part 5: The Execution Engine & Portfolio State Machine. Bridging the gap to live execution with Day-1 stop-loss immunity, rank-based pending order allocation, and database garbage collection for zombie orders. Part 1 will be posted shortly. Keep an eye out for it, and if you're interested let me know in the comments which part of the stack you are most interested in diving into.
-
Pradeep Jindal reacted on thisPradeep Jindal reacted on thisהיינו היום בים. וחשבתי על החג הזה ומה הוא מסמל. התחלה חדשה. חשבון נפש. התחדשות. שאלתי את עצמי מה מאחלים לעצמם כל האנשים בים? הסבא בן ה80 שקורא עיתון, הזוג הצעיר שנרגש להיכנס יחד למים, ההורים העייפים שהביאו אוכל כדי לחסוך את הערב בבית, הברמנית שהגישה לי בירה. הם רוצים התחדשות? טוב להם? איך עברה עליהם השנה האחרונה? ואז חזרתי אל עצמי, ולמשפחה הקטנה שלנו - וייחלתי לנו ימים של נחת, ושקט, עם הנהגה חדשה ואחדות אמיתית. קצת פחות לחות. קצת יותר אוויר. שנה אחת בלי מלחמה אחת. שעמום. זמן משפחה. עבודה קשה. סיפוק. שנה טובה.
-
Pradeep Jindal reacted on thisPradeep Jindal reacted on thisHonored to be recognized as DevSecOps Leader of the Year at AppSec Live 2026. Grateful for the recognition, and for the work we’ve been doing at Scapia to make security and reliability foundational to how we build and operate. This recognition is a reflection of the people and the work behind it. It was a pleasure to exchange perspectives with peers and industry leaders on the evolving AppSec and DevSecOps landscape. Thank you to the organizers and sponsors for bringing the community together and creating space for meaningful conversations. #DevSecOps #AppSec #Cybersecurity #EngineeringLeadership #SRE #EngineeringExcellence #Reliability #CloudSecurity #DevOps #TechLeadership #TechnophilesIndia #AppSecLive2026
-
Pradeep Jindal reacted on thisPradeep Jindal reacted on thisWhen I first found out my wife performed surgery at 4am, 28 hours into a 36-hour shift, I thought: ‘this is insane.’ We started Konko AI so she wouldn’t have to. Today we're announcing our $6 million raise led by Hi Ventures to give doctors their lives back and let them serve patients the way they always wanted to. It took us years to get here. Five years ago, my wife and I started a private clinic. Our thinking: if we owned a business she’d have more control over her schedule. Wrong. The business succeeded. We grew to thousands of patients and became one of the top-rated OBGYN practices in Latin America. But it didn’t really work. She was still pulling 100-hour weeks. We realized the problem isn’t owning a business vs. not. The problem is supply and demand. Doctor supply is finite. Demand for health is infinite. So we teamed up with Michael Haddad, my friend, HBS classmate and two-time technical founder who built mass-scale AI systems at C3.ai (pre- and post-IPO) and other venture-backed startups, to make healthcare infinite. AI makes intelligence abundant and cheap. But that’s not enough to deliver infinite healthcare. You need healthcare specific automation and full patient context. Konko AI is the automation and context layer for global healthcare. We’re starting with Latin America, a region of 660 million people with a third fewer doctors per capita than the OECD average. 🇲🇽 🇧🇷 🇨🇴 🇨🇷 🇦🇷 🇻🇪 A year after launch we’ve: > Saved years’ worth of doctor and clinical staff time. > Handled over 2 million patient interactions. > 75% of these patient interactions were handled autonomously by AI with 4.5/5 patient satisfaction. 2 million patient interactions in. 5.6 billion patients to go. More in the comments 👇
-
Pradeep Jindal liked thisPradeep Jindal liked thisExcited to share a blog I recently worked on at Yugabyte! We explored how popular PostgreSQL CDC tools like Sequin, PeerDB, and Supabase ETL work with YugabyteDB, and what it takes to stream changes from a distributed PostgreSQL-compatible database. A great learning experience diving deeper into CDC, logical replication, and distributed databases. Check it out! #YugabyteDB #PostgreSQL #CDC #DistributedSystemsHow PostgreSQL CDC Tools Work on YugabyteDB | YugabyteHow PostgreSQL CDC Tools Work on YugabyteDB | Yugabyte
-
Pradeep Jindal liked thisPradeep Jindal liked thisExcited to be speaking at The Fifth Elephant on 31 July 2026 at the NIMHANS Convention Centre, Bengaluru. I’ll be presenting a live demo: Beyond RAG: Turning Financial PDFs into Queryable Tables We’ll explore how agentic AI can move beyond simply retrieving answers from documents—transforming complex, unstructured financial PDFs into reliable, structured tables that can be queried and analysed using OrcaSheets. If you’re attending The Fifth Elephant, I’d love to connect and exchange ideas on document intelligence, AI agents, and the future of data analysis. See you there! #TheFifthElephant #AgenticAI #ArtificialIntelligence #DocumentIntelligence #FinancialData #DataAnalytics DataOrc OrcaSheets
-
Pradeep Jindal liked thisPradeep Jindal liked thisI’m spending the weekend at a beautiful, secluded hotel in the desert. This morning, I noticed an employee whose entire role appears to be making sure guests don’t leave a towel and a bag on a pool sunbed, go to breakfast, and 'reserve' it for the next two hours. A deeply Israeli problem. And apparently, now, a real job. The hotel identified a recurring user behavior that creates frustration for everyone else. Signs didn’t solve it. Social norms definitely didn’t solve it. So they added an operational layer to enforce the intended experience. No AI. No algorithm. No 'smart poolside optimization platform.' Just one person protecting the sunbed supply. Sometimes the best product solution isn’t a new feature.. It’s accepting that your users will behave exactly like your users, and, designing the system accordingly. (:
Experience & Education
Projects
-
Databus (Highly Available, Highly Scalable data pipeline for INMOBI)
-
Worked on various components of the Databus project, including scribe (the backbone of Databus data flow), thrift (the backbone of scribe), Operability for the whole project. The work includes:
1. Instrumenting/Augmenting scribe to expose more useful metrics.
2. Augmenting scribe to behave correctly w.r.t operability e.g. signal handling, *really* graceful shutdown etc.
3. Patching the thrift Framed Transport to fix the duplicate messages problem. ( https://github.com/praddy/thrift…Worked on various components of the Databus project, including scribe (the backbone of Databus data flow), thrift (the backbone of scribe), Operability for the whole project. The work includes:
1. Instrumenting/Augmenting scribe to expose more useful metrics.
2. Augmenting scribe to behave correctly w.r.t operability e.g. signal handling, *really* graceful shutdown etc.
3. Patching the thrift Framed Transport to fix the duplicate messages problem. ( https://github.com/praddy/thrift )
4. Augmenting the producer side client to take "acks" into account.
5. Working with the Ops team and suggesting optimal ways to monitor the project in production.Other creatorsSee project -
Fraud Detection Service
-
Kickstarted the concept of a separated dedicated offline Fraud Detection Service to remove the burden off of the online front-end servers. A very capable engineering team took it further and got it through to production. This helped in reducing the response times of the front-end servers and by removing the online storage required we were able scale the front-end servers out across regions.
Other creators -
imetalib (domain specific ORM with an adapter for Redis)
-
As the name suggests, it is an ORM (domain specific, written in pure Ruby) for metadata services (about production servers) consisting of following entities:
1. Node Facts (cpu, memory, free memory etc.)
2. Node Metadata (cluster, partition, product etc.)
3. Node Volatile Facts (facts with TTL e.g. for use in auto deployment, alerting etc.)
Currently a UI layer and a command-line is being written around it which will facilitate inventory management and a metadata store for…As the name suggests, it is an ORM (domain specific, written in pure Ruby) for metadata services (about production servers) consisting of following entities:
1. Node Facts (cpu, memory, free memory etc.)
2. Node Metadata (cluster, partition, product etc.)
3. Node Volatile Facts (facts with TTL e.g. for use in auto deployment, alerting etc.)
Currently a UI layer and a command-line is being written around it which will facilitate inventory management and a metadata store for use with monitoring, auto deployment tools/frameworks. -
iMon (generic, scalable, pluggable, decoupled monitoring framework)
-
Designed and Architected iMon. And also helped in implementing the design. Written in Perl + C, iMon maps the monitoring process back to basics of computing:
Producing Data -> Collecting Data -> Processing Data -> Rendering Processed Data (or Producing Data to be consumed by next round)
And all this in a very generic and pluggable way. On top of iMon we wrote Metrics Aggregator (as a Processor) which supports average, moving average, histograms, 90th percentile etc. as…Designed and Architected iMon. And also helped in implementing the design. Written in Perl + C, iMon maps the monitoring process back to basics of computing:
Producing Data -> Collecting Data -> Processing Data -> Rendering Processed Data (or Producing Data to be consumed by next round)
And all this in a very generic and pluggable way. On top of iMon we wrote Metrics Aggregator (as a Processor) which supports average, moving average, histograms, 90th percentile etc. as aggregation functions, more can be added easily. We also wrote Tagger (as a Processor) which includes a power rule engine which can be used to tag the messages and then act upon tags e.g. alert, route etc.Other creatorsSee project -
Augmented Puppet for Metadata features
-
Added a new Puppet Domain Specific Language construct to populate metadata about nodes while leveraging the full feature set of the Puppet Domain Specific Language. Now puppet had a way to tag or add metadata (like cluster info, rotation info, alerting info etc.) to nodes which then can be used by various tools/frameworks to achieve goals like automated deployment, seamless migration, smart alerting/monitoring etc. Its currently in use in production.
-
Apex DOTS Pipeline implementation and Migration of Services.
-
Lead the effort of implementing a new DOTS Pipeline across several regions/datacentres and migration of services from older pipeline instance(s). DOTS (DistrO Transformation Server) Pipeline is the Highly Available & Highly Scalable backbone for all the advertisers' and publishers' updates from the UI layer.
Other creators -
Contributed and enhanced various in-house frameworks/tools
-
These included:
1. Monitoring Framework (pure Perl)
2. Generic Multiplexed TCP/SSL library (Perl)
3. Coding standard for operations team.
-
Virtual Clusterization of Bittorrent Peer Network
-
The problem was "How to keep the benefits of Bittorrent while removing the network noise created by it?" because Bittorrent started to choke our switches and Load Balancers. Designed and implemented a peerlist filter mechanism inside the Bittorrent tracker while keeping following things in mind:
1. No changes to tracker protocol to ensure any client can be used thereafter.
2. No changes to the peer-wire protocol to ensure cross-client operability.
3. Changes to minimal components…The problem was "How to keep the benefits of Bittorrent while removing the network noise created by it?" because Bittorrent started to choke our switches and Load Balancers. Designed and implemented a peerlist filter mechanism inside the Bittorrent tracker while keeping following things in mind:
1. No changes to tracker protocol to ensure any client can be used thereafter.
2. No changes to the peer-wire protocol to ensure cross-client operability.
3. Changes to minimal components in the system.
On top of the framework, implemented a subnet peerlist filter to virtually partition the large Bittorrent cluster in subnet based networks. Switches and Load Balancers doing around 70-80% cpu came down to 5-10% -
Operability for Dispaly Advertising Products
-
Reviewed all the backend products for Display Advertising Group and suggested:
1. Instrumentation Points
2. Way to Monitor
3. Architecture Enhancements
4. Ways to remove SPOFs
to make sure all the products are highly operable. -
Lead the Indian Chapter of APG Operations' Platform Group
-
Lead the Indian chapter of the APG (Advertiser Publisher Group) Operations' Platform Group. Included reviewing, designing, implementing and project managing generic ideas,frameworks and tools.
-
ERAF (Email Reporting and Assertion Framework)
-
As the sole member of the QA Automation team in India, I designed and implemented the ERAF. ERAF was essentially a Domain Specific Language with a platform for receiving and storing emails + reporting features written in pure Python. It was intended to QA the gateway level Email Security Appliance from Secure Computing (later acquired by McAfee). It helped the QA team to define workflows of test cases, which made the QA process very effective and time efficient by a factor of 10.
Other creators -
Swiss Army Knife for Emails
-
Wrote a very powerful email tool for the command-line in pure Python. Again for the QA team this tool cloud:
1. Dissect stored (in any storage format) or received emails.
2. Forge thousands of different emails per second and then either send them to single/multiple targets or store them in the desired storage format for later use.
3. OR a combination of all the above just from a single simple command-line.
This helped the QA team to conduct lengthy complex tests in minutes…Wrote a very powerful email tool for the command-line in pure Python. Again for the QA team this tool cloud:
1. Dissect stored (in any storage format) or received emails.
2. Forge thousands of different emails per second and then either send them to single/multiple targets or store them in the desired storage format for later use.
3. OR a combination of all the above just from a single simple command-line.
This helped the QA team to conduct lengthy complex tests in minutes rather than days. QA team went time efficient by a factor of 10. -
Cybernetra Mail Server
-
See projectAgain putting control in the Enterprises' hands, the ones who care about it. Architected Cybernetra MailServer to be based on various successful open source technologies like postfix, spamassassin, LDAP + a very powerful properietary rule engine which plugged into various stages of email flow throughout the system. The rule engine introduced the "Policy Based Mailing" concept which put the Enterprises' right in control.
-
Cybernetra Internet Management System
-
Gateway level Internet Management System for Enterprises. Rearchitected the whole stack to perform better in terms of throughput as the number of per user management rules on server side grow larger. This included, C development, Analyzing the Linux Kernel network stack, Analyzing layer 3 and above protocols, Analyzing various classfull and classless bandwidth management algorithms, optimizing performance of Squid and intergrating Squid with MySQL. This role included managing a team of varying…
Gateway level Internet Management System for Enterprises. Rearchitected the whole stack to perform better in terms of throughput as the number of per user management rules on server side grow larger. This included, C development, Analyzing the Linux Kernel network stack, Analyzing layer 3 and above protocols, Analyzing various classfull and classless bandwidth management algorithms, optimizing performance of Squid and intergrating Squid with MySQL. This role included managing a team of varying skillsets.
Other creatorsSee project
Recommendations received
9 people have recommended Pradeep
Join now to viewOther similar profiles
Explore more posts
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore MoreOthers named Pradeep Jindal
62 others named Pradeep Jindal are on LinkedIn
See others named Pradeep Jindal