Skip to content
View tkgo1599-max's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report tkgo1599-max

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
tkgo1599-max/README.md

Gowtham Thirukkovalluri — AI/ML Engineer

LinkedIn Email Location Profile views

I build the unglamorous machinery that makes ML trustworthy in production: eval harnesses that fail bad prompt changes in CI, RAG pipelines with measured recall, drift detectors with defensible thresholds, and promotion gates that make "ship it" a reviewed, reversible decision. If it doesn't have a number and a rollback path, it isn't done.

⏳ 4+

Years in prod ML
OpenAI · Amazon · Research

📦 10

Engineering case studies
tested · benchmarked · documented

📡 12+

LLM endpoints monitored
p95, drift & hallucination rates

🏆 2

Certifications
GenAI w/ LLMs · Claude on Bedrock

🎛️ Experience levels

LaneSignalLevel
LLMs · RAG · Evals ▰▰▰▰▰▰▰▰▰▱ daily driver
MLOps · Serving · CI/CD ▰▰▰▰▰▰▰▰▰▱ daily driver
Forecasting · Recommenders ▰▰▰▰▰▰▰▰▱▱ production proven
Streaming · Data Eng ▰▰▰▰▰▰▰▱▱▱ production proven
Monitoring · Observability ▰▰▰▰▰▰▰▰▱▱ production proven

🔭 Current focus

  • 🧪 LLM evaluation & prompt-regression testing in CI - harnesses that caught 15+ regressions before rollout; open-sourced the pattern → prompt-regression-gate
  • ⚡ Low-latency inference - REST/gRPC microservices at 99.9% uptime; batching & shedding mechanics → low-latency-inference-gateway
  • 📉 Drift detection that on-call can trust - calibrated thresholds over folklore constants → drift-sentinel
  • 💸 GPU & token FinOps - 28% fewer GPU-hours per fine-tuning run at work; router experiments → llm-cost-router

🧱 Building blocks

Tech stack icons

LangChain Hugging Face MLflow SageMaker Airflow Spark Pinecone ChromaDB XGBoost / LightGBM

📌 Featured builds

Repo What it proves
🚦 prompt-regression-gate CI gate for LLM changes - 1,000-case eval suite in 2.2s, zero API calls
🔍 rag-quality-bench Chunk size moved MRR 0.43 → 0.91; retriever choice moved it ≤0.03
📉 drift-sentinel Calibrated PSI thresholds catch drift the folklore 0.25 rule misses
🌊 stream-alert-engine Incident flagged 20s after onset, 0 false positives @ 800k events/s
📈 forecast-backtest-kit Leak-proof backtests - corruption probe as a CI-enforced guarantee
⚡ low-latency-inference-gateway Micro-batching: +54% throughput, p99 2289→915ms under load
🧬 lora-efficiency-lab LoRA from the math - 0.4% of params, 99.7% accuracy, and when it fails
🚢 model-promotion-gate Champion/challenger promotion as one reviewed, reversible alias move

🔀 Pivot point

Started as a Data Scientist at Amazon, forecasting demand for 200+ SKU categories and shipping recommenders that lifted CTR 14% - classical ML with a business number attached to every model. Then the LLM wave hit, and the interesting problem moved: it wasn't training models anymore, it was making generative systems behave in production - evaluation, drift, latency, cost. So I pivoted through an AI research assistantship (RAG systems, recall@10 of 0.91 across 100K+ embeddings) into LLM platform engineering, where my classical-ML instincts (baselines, backtests, error budgets) turned out to be exactly what LLMOps was missing. The through-line: I don't ship vibes. I ship measured systems.

🎮 Outside the code

🏏 Cricket on weekends  ·  ♟️ Chess puzzles with my coffee  ·  🎬 Tollywood film nights  ·  🍳 Perfecting one biryani recipe per month

🤝 Quick connect

LinkedIn Email GitHub

Pinned Loading

  1. prompt-regression-gate prompt-regression-gate Public

    CI gate that blocks LLM prompt and model changes which regress on a golden eval suite. Three deterministic checks (token-F1, grounding, JSON Schema) against git-committed baselines; gates 1,000 cas…

    Python

  2. rag-quality-bench rag-quality-bench Public

    Retrieval-quality harness for RAG: recall@10 and MRR against labeled queries across chunkers and retrievers. Chunk size moved MRR 0.43 to 0.91 while swapping retrievers moved it under 0.03. SQLite …

    Python

  3. drift-sentinel drift-sentinel Public

    Feature drift detector with alert thresholds calibrated from your own reference data (p99 of no-drift PSI via self-splits), not folklore constants. Flags a PSI-0.18 drift the 0.25 rule misses. Fast…

    Python

  4. forecast-backtest-kit forecast-backtest-kit Public

    Rolling-origin demand-forecasting backtest with a provably leak-free feature pipeline: corrupting every post-origin actual changes zero predictions (CI-enforced). LightGBM vs honest baselines, WAPE…

    Python

  5. low-latency-inference-gateway low-latency-inference-gateway Public

    Model-serving gateway with dynamic micro-batching (fill-to-32 or 4ms deadline) and load shedding: +54% throughput and p99 cut from 2,289ms to 915ms at 64-way concurrency vs naive serving. Documents…

    Python

  6. model-promotion-gate model-promotion-gate Public

    Champion/challenger promotion gate on the MLflow registry: challengers earn the champion alias by beating the incumbent on a frozen eval set within a latency budget, or stay blocked with the failin…

    Python