BALAJI K
AI/ML ENGINEER building LLMs, RAG & agentic systems from first principles
SCROLL
I design and ship agentic AI systems end to end — model internals, retrieval, and multi-agent orchestration — and I hold every one of them to the same bar: measured, not assumed. That means hallucination rate and citation faithfulness benchmarked across quantization levels, guardrails and access scoping built into an agent from day one instead of bolted on after, and cost/latency tradeoffs backed by numbers instead of guesses. I'm comfortable reading an unfamiliar problem fast and designing the architecture that actually fits it, rather than defaulting to "put an agent on everything."
A four-lane fraud-detection architecture for UPI/IMPS/card payments, built after routing every transaction through an LLM agent proved too slow and too expensive at bank scale. A gradient-boosting model scores every transaction in a hot lane; an LLM agent investigates only the ambiguous ~1% in a gray-zone lane; a verification lane confirms high-risk payments. A privacy gateway pseudonymizes identifiers before anything reaches the LLM, and a compact encoder compresses each case from ~457 tokens to ~25 before it hits the agent. On held-out synthetic data: 0.76 precision / 0.94 recall at a fixed 0.1% false-positive rate, with 98.35% of transactions never leaving the hot lane.
An open-source (MIT) middleware layer for high-volume, repetitive agentic workloads — fraud or compliance pass/fail checks, for example. It sits between an application and its LLM agent, replacing verbose JSON payloads and raw identifiers with short aliases and derived signals on the input side, and compact reason codes expanded locally on the output side. In internal testing: ~50–70% fewer input tokens and ~60–75% fewer output tokens per request, with no loss of decision information at the protocol layer.
An end-to-end multi-agent AI system, built on Azure AI Foundry, that runs the operations of a textile business — orchestrated agents sitting over a web and database layer, deployed as a live production product rather than a prototype.
An enterprise agentic assistant wrapping GPT-4o-mini in a production-style pipeline designed around guardrails from the start: every request is access-scoped before it runs, retrieval grounds answers in company data instead of the model's memory, and six tool integrations (email, calendar, HR, docs, approvals) execute only after human-in-the-loop approval — the whole flow traced and gated by an eval suite in CI/CD. The little mascot guiding this site is named after her.
A fully local, Perplexity-style search-and-answer engine running entirely on an 8GB M2 Mac. A 34M-parameter transformer, trained from scratch in PyTorch with a custom chunk-isolated attention mechanism for multi-source reading (perplexity 4.9 on held-out data), sits over a BraveSearch RAG pipeline alongside a quantized Qwen2.5-3B production layer. Redesigning the model's output format lifted grounded-QA accuracy from 10% to 78% on held-out evaluation, and hallucination/citation faithfulness is benchmarked across FP16, Q8, and Q4 quantization.
A recall-optimized ResNet-18 for lymph node metastasis detection on H&E histopathology patches (PatchCamelyon), reaching 0.977 ROC-AUC and 0.950 recall — on par with the published PCam benchmark range. Its decision threshold is auto-tuned to ≥95% sensitivity to match clinical screening standards, with Grad-CAM explainability and a Gradio inference UI.
Open to AI engineer and AI research roles, and Forward Deployed Engineer positions.
Let's talk↗