Hi, I'm Krishna Koushik Unnam.
I train and ship large-scale AI systems that run in production.
AI/ML Engineer with 5+ years designing, training, and deploying foundation models, multimodal LLMs, RAG, RLHF, and enterprise ML — from GPU clusters to cloud inference serving millions of users.
Currently building: multimodal foundation model training, Mixture-of-Experts architectures, RLHF/DPO alignment, RAG systems, and agentic AI platforms at Meta.
Focus
Foundation Models, RAG, MLOps
Based in
Menlo Park, CA
About
Research-grade models, production-grade
engineering.
I'm an AI/ML Engineer with 5+ years of experience designing, training, and deploying large-scale AI systems — including foundation models, multimodal LLMs, and enterprise machine learning solutions. Experienced in distributed model training, Generative AI, RAG, RLHF, MLOps, and scalable inference across GPU clusters and cloud platforms, delivering production-ready AI for millions of users.
At Meta, I lead multimodal training pipelines, Mixture-of-Experts optimization, SFT/RLHF/DPO alignment, RAG systems, and inference platforms that serve hundreds of millions of requests daily. Before that at Accenture, I shipped industrial predictive maintenance, computer vision inspection, digital twins, and cloud-native MLOps from petabyte-scale sensor data to real-time scoring APIs.
I'm passionate about building intelligent systems that combine cutting-edge research with robust engineering — measurable quality gains, governed LLMOps, and inference that stays reliable and cost-efficient at scale.
Tech Stack
Technologies I build with
Languages, frameworks, training stacks, data platforms, and MLOps tooling I use to design, train, evaluate, and serve large-scale AI systems.
Core
Also Used
Core
Also Used
Core
Also Used
Core
Also Used
Core
Also Used
Core
Also Used
Core
Also Used
Core
Also Used
Core
Also Used
Enterprise AI platforms I helped
build
End-to-end ownership across foundation model training, alignment, RAG, industrial ML, and production inference — shipped outcomes at scale.
Meta — Multimodal Foundation Models
AI/ML Engineer building large-scale multimodal foundation model training, MoE optimization, RLHF/DPO alignment, RAG, agentic platforms, and inference serving more than 600 million global users daily across AWS, Azure, GCP, and hyperscale private data centers.
Accenture — Industrial AI & MLOps
Machine Learning Engineer delivering predictive maintenance, computer vision, digital twins, and cloud-native ML platforms for manufacturing — from billions of sensor records through real-time scoring APIs.
Selected Projects
Selected work worth exploring
Each project includes the engineering decision that mattered most — not just what was built, but why it was built that way.
Production multi-provider LLM inference gateway with score-based routing, pgvector semantic caching, JWT tenant auth, Redis token-bucket rate limiting, circuit-breaker fallback, SSE streaming, and Prometheus/Grafana observability — load-tested at 0% failures, 18.43 req/s, 567 ms p95, and 99.97% cache hit rate across 3,885 requests at 100 VUs.
Key Insight
Semantic cache writes run in a background goroutine so they add zero latency to the caller — the cost lever only works if cache population never blocks the hot path.
Tech Stack
Stateful issue-to-pull-request workflow — repository research, minimal-fix planning, commit-pinned patch generation, Docker-isolated validation, structured retry feedback, and human escalation. Validated on real Click issues: 1,136–1,436 tests passing, clean Ruff/Mypy, zero retries, ~2 min runtime, ~$0.35–$0.36 model cost.
Key Insight
Unified diffs fail on pinned historical SHAs because models invent context lines — search/replace blocks fetched at base_commit_sha sidestep that entire class of apply failures.
Tech Stack
Explainable enterprise RAG platform — extracts and chunks PDFs, stores document-scoped vectors in Pinecone, and serves context-grounded answers via FastAPI + React. Document/chunk metadata filtering, retrieved-source evidence, empty-retrieval handling, and LlamaIndex relevancy evaluation reduce cross-document contamination.
Key Insight
The core design principle: every answer must be traceable to a source passage. Surfacing the exact chunk shifts user trust from 'I hope this is right' to 'I can verify this is right.'
Tech Stack
Benchmarked LoRA vs Adapter fine-tuning across 3 cardiac foundation models on 65K+ ECG segments. Wav2Vec2+LoRA achieved best balanced performance (F1: 0.589, AUC: 0.620) with 60% fewer trainable parameters than full fine-tuning.
Key Insight
Audio models have no token embeddings — injecting LoRA required patching enable_input_require_grads to a no-op, then using mask-aware mean pooling across downsampled sequence lengths.
Tech Stack
ResearchFlow decomposes any query into parallel sub-tasks, retrieves and ranks web evidence, and delivers a cited, self-verified answer in under 30 seconds — using LangGraph's stateful graph to enable retry loops that a linear chain couldn't support.
Key Insight
Chose LangGraph over a simpler LangChain chain specifically because stateful graphs allow conditional retry edges when source relevance scores fall below a confidence threshold.
Tech Stack
ContextFlow runs entirely on-device using Ollama — 100% local inference, zero external calls, no data leaves the browser. Summarises pages, explains highlighted text, and answers context-aware questions with sub-second response times on consumer hardware.
Key Insight
The hardest part wasn't the LLM integration — it was building a streaming response protocol between the Chrome extension content script and the background service worker without a shared DOM.
Tech Stack
Workforce management system for 500+ transport authority employees — Spring Boot REST APIs with JWT/RBAC across 5 roles, optimistic locking on shift-assignment records, duty assignment and leave approval workflows. Scheduling overhead reduced 40%.
Key Insight
Optimistic locking prevents a subtle race condition where two admins simultaneously assign the same employee to overlapping shifts — catching the conflict at the database level rather than application level.
Tech Stack
17.1% FID improvement (290.19→240.48) with 85% reduction in training time. Enhanced conditional UNet with 16×16 multi-head self-attention, DDIM sampling extended 50→100 steps. CFG ablation: higher guidance = better fidelity, lower = better diversity.
Key Insight
Fine-tuning outperforms training from scratch for conditional generation — the pretrained UNet already encodes useful low-level structure that would take thousands of steps to relearn.
Tech Stack
Inference-time best-of-n selection using CLAP (text-audio alignment) scoring — ~40% CLAPScore improvement at n=5 with no training cost. Includes spectrogram comparisons and cost-benefit analysis showing clear diminishing returns beyond ~7 samples.
Key Insight
Best-of-n selection shifts quality from 'average output' to 'peak capability of the model' — the real challenge is that CLAP scoring itself costs ~2s per candidate, so the cost-benefit curve peaks around n=5.
Tech Stack
Unpaired image-to-image translation between Monet paintings and photographs. Ablation studies on cycle-consistency (λ_cyc ∈ {5,10,20}) and identity loss. λ_cyc=10 produced the most visually balanced results with stable colour preservation.
Key Insight
λ_cyc=10 is the sweet spot: lower values allow stylistic flexibility but weaker reconstruction fidelity; higher values over-constrain the generator and limit the range of stylization.
Tech Stack
Full-stack alternate implementation using ASP.NET Core + Entity Framework Core. JWT/HMAC-SHA256 auth, Redis-backed token blacklisting for stateless auth with revocation, rate limiting middleware, layered architecture with OpenAPI docs.
Key Insight
Redis-backed JWT blacklisting enables stateless auth with revocation — solving the logout problem without session state by keeping a small TTL-expiring blacklist that mirrors token lifetime.
Tech Stack
Enterprise Experience At
Experience
Where I've worked
Company-backed AI/ML engineering across foundation models and industrial machine learning — with exact dates, locations, and scope.
AI/ML Engineer
Meta · Menlo Park, CA
Develop large-scale multimodal foundation model training, MoE architectures, RLHF/DPO alignment, RAG systems, evaluation frameworks, MLOps, and high-throughput inference (vLLM, TensorRT-LLM) — serving more than 600 million global users daily across AWS, Azure, GCP, and hyperscale private data centers.
Machine Learning Engineer
Accenture · India
Built industrial ML spanning predictive maintenance on 15,000+ IoT assets, computer vision inspection (YOLO), digital twins, Spark/Databricks platforms, and AWS SageMaker MLOps — delivering real-time scoring APIs with Java/Spring Boot and FastAPI, plus Power BI stakeholder dashboards.
Education
Academic foundation
Formal training in computer science and engineering, backed by cloud and ML certifications.
Master of Science in Computer Science & Engineering
University of South Florida
Certifications
Credentials that back the work
AWS Certified Generative AI Developer – Professional (AIP-C01)
Databricks Certified Machine Learning Professional
Certified Kubernetes Application Developer (CKAD)
Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103)
Contact
Let's connect
Open to AI/ML, foundation model, and GenAI roles. Whether you have an opportunity, a collaboration in mind, or just want to say hi — my inbox is open.