# Neerudi Sai Vikas

> Applied AI Engineer in Hyderabad, India. I turn ambiguous problems into AI systems with real impact. I work directly with clients, from the first conversation to production: voice agents, document and search tools, and financial systems where every number has to be right.

- Name: Neerudi Sai Vikas
- Role: Applied AI Engineer
- Location: Hyderabad, India · Open to remote and relocation
- Current: Backend & AI Systems Engineer at TechRAQ (Apr 2026 – Present)
- Previously: AI Systems Engineer & Applied ML Researcher, Independent AI Systems Engineering & R&D (Jul 2022 – Apr 2026)
- Education: B.Tech, Computer Science & Engineering, Indian Institute of Technology (IIT) Indore (Graduated Jul 2022)
- Looking for: forward-deployed and applied AI engineering roles
- Email: neerudisaivikas@gmail.com
- LinkedIn: https://www.linkedin.com/in/neerudi-sai-vikas-721349226/
- GitHub: https://github.com/kee1bo
- Resume: https://saivikas.work/assets/resume.pdf

## Selected projects

- [Course builder for an EdTech platform](https://saivikas.work/projects/autonomous-course-engine): Turns a company's raw documents into complete training courses: lessons, slides, quizzes and voiceovers. Each client's data stays fully separate. Cut manual authoring time by about 70%.
- [Phone agent for a US logistics company](https://saivikas.work/projects/realtime-voice-cx-agent): Answers live customer calls about orders and warehouse status. Replies start in under 350 ms, and callers can only ever reach their own data.
- [Reservoir: fund reconciliation you can audit](https://saivikas.work/projects/reservoir-evergreen-analytics): The AI reads messy fund reports; plain code does all the math. Every number traces back to the exact cell it came from. Passed all 43 audit checks on a synthetic quarter of fund data.
- [A language model, built from scratch](https://saivikas.work/projects/first-principles-slm-compression): Wrote and trained a 21M-parameter GPT-style model in PyTorch, tokenizer included. Now using it to test ways to make small models cheaper to run.

## Writing

- [Two ways to make small language models cheaper](https://saivikas.work/blog/shannon-entropy-patching-and-csa2-kv-compression): Spend compute only where text is hard to predict, and let layers share memory instead of each keeping their own.
- [Let the model read. Let code do the math.](https://saivikas.work/blog/bitemporal-financial-reconciliation-neuro-symbolic-agents): Why an AI system for fund operations should never do arithmetic itself, and how Reservoir keeps reading evidence separate from computing numbers.
- [From 2.5 seconds of silence to replies in under 350 ms](https://saivikas.work/blog/sub-350ms-cascaded-telephony-rag-zero-trust-authz): How we made a phone agent fast enough for natural conversation, and kept each caller locked to their own data.

## Optional

- [Altium: research and buyback planning for funds](https://saivikas.work/projects/altium-evergreen-funds): Helps wealth advisors keep track of the funds they hold, and helps fund teams plan quarterly buybacks. Forecasts come from 2,000 simulated scenarios, and every claim cites its source.
- [A trading engine that adapts to the market](https://saivikas.work/projects/hierarchical-rl-quant-engine): First works out which of three market regimes it is in, then lets a reinforcement-learning agent decide what to do in that regime. Tested on historical data with no look-ahead.
- [Research assistant over 100k+ papers](https://saivikas.work/projects/scholar-hybrid-rag): Combines keyword and meaning-based search across arXiv papers and keeps only the useful passages in memory. Recall@10 went from 0.62 to 0.89, at 40% lower token cost.
- [Cairn: private credit monitoring](https://saivikas.work/projects/cairn-private-credit-monitor): Reads loan documents and financial statements whatever their layout, pulls out the key figures with citations, and computes the numbers a credit team checks every quarter.
- [VoiceOps: reviewing sales calls](https://saivikas.work/projects/voiceops-call-intelligence): Scores sales calls against a rubric and remembers context across calls. Caching cut a slow database read from 720 ms to 2 ms; routing work to the right model cut chat replies from 46 s to 11 s.
- [Sorting invoices by layout](https://saivikas.work/projects/invoice-layout-clustering-cv): Before modern vision models: grouped a large set of raw invoices into 5 layout types without labels, then used a tuned extractor for each type to read fields and tables.
- [Automatic billing for truck parking](https://saivikas.work/projects/commercial-truck-parking-cv): Cameras spot trucks entering and leaving a San Francisco parking yard, track how long each one stays, and bill for it automatically.
- [Spotting tampered documents](https://saivikas.work/projects/enterprise-doc-forensics): Flags edited or forged documents during customer onboarding by inspecting the image for signs of editing and cross-checking fields against each other.
- [A voice assistant in the Linux top bar](https://saivikas.work/projects/island-wayland-voice-runtime): A GNOME extension that listens, transcribes speech on the laptop, hands it to an AI agent and speaks the reply. Speech in and out runs locally, and replies start in under 0.6 s.
- [Full profile in one file](https://saivikas.work/llms-full.txt): experience, every project in detail, and all writing