Work Experience Toolkit About Contact
Software Engineer, IBM MS, University of Pittsburgh · United States

$ I build backend systems that have to be right under load.

At IBM: the Guardium Sniffer, a C++20 engine that decodes live Oracle and MySQL traffic.

On my own time: AI-agent infrastructure, from MCP servers to a terminal coding agent.

See the work

~/work $ ls --featured

Selected work

Three systems I designed and built end to end in 2026, a RAG assistant for Chicago parents, then my AI engineering projects, the systems behind my resume, and everything else worth a look. Every number here comes from the repo or my own resume.

year
2026
stack
Python 3.12, FastAPI, PostgreSQL, Redis, TypeScript MCP server, Docker, GitHub Actions
size
~17,200 lines Python · 2,350 TypeScript · 17 test files

Sentinel Execution MCP

A trading control plane that an AI agent drives through the Model Context Protocol. The agent asks; the engine decides.

Sentinel architecture: an AI agent calls a TypeScript MCP server, which forwards to a FastAPI engine that runs regime classification, an 8-voter committee, a 14-check risk firewall, sizing, broker execution and audit, backed by PostgreSQL and Redis. packages/engine · FastAPI AI agentany MCP client packages/mcpTypeScript · 40+ toolscircuit breaker:opens after 5 failures RegimeATR·ADX·Hurst… Committee8 voters, 3 vetoes Risk firewall14 risk checks Sizing1% risk / trade Brokerpaper · Alpaca Audit journalexplains any trade PostgreSQLorders · positions · audit Rediskill switches · rate limits HTTP
fig. 1 — request path for a single ordercircuit breaker: closed → open → half-open, retry after 30 s

Problem

Letting a language model place trades is only safe if the model never holds trading logic. Every decision has to be explainable, reversible and gated.

Approach

  • Thin MCP server; all logic lives in the engine.
  • Regime classifier scores six strategies per market regime.
  • Eight-filter committee: ≥0.65 approves, 0.45–0.65 goes to a human.

Decisions

  • Fail fast: if the engine is down, every tool errors with no partial execution.
  • Strategies move draft → research → backtest → paper → live, each gate with criteria; paper to live needs a human.
  • API keys stored only as SHA-256 hashes.
40+MCP tools in 9 groups
14pre-trade risk checks
6strategies + walk-forward backtests
year
2026
stack
TypeScript, Node 20, pnpm monorepo, Ink, Bun-compiled binaries
ships
macOS arm64 / x64, Linux x64, Windows
source
Core repo private (MIT).
github / forge-releases

Forge

A terminal-native coding agent, built as a reusable SDK rather than a closed tool.

Forge packages: the CLI drives @forge/core, which runs the agent loop and permission gate and uses @forge/models, @forge/tools and @forge/plugins; models connect to OpenAI, Anthropic and Ollama through a fallback chain. @forge/cliInk REPL · /slash cmds @forge/coreagent loop · event busmiddleware pipelinepermission gate + trust levels @forge/modelsprovider adapters @forge/toolsread·edit·shell·grep·MCP @forge/pluginsdefinePlugin, hooks ProvidersOpenAI · AnthropicOllama FallbackChain
fig. 2 — five packages, one agent loopany OpenAI-compatible endpoint works

Problem

Most coding agents are closed products. I wanted the agent loop, tools and provider layer as parts someone else could build on.

Approach

  • Split into core, models, tools, plugins and CLI packages.
  • Streaming end to end: SSE parsing into async generators.
  • Reads CLAUDE.md, supports --continue, runs dev servers in the background.

Decisions

  • Every tool call passes a permission gate with configurable trust levels.
  • Middleware wraps model calls, so logging, caching and token accounting are pluggable.
  • Blocks long-running server commands that would freeze the agent.
515security + adversarial tests
5packages
4platform binaries per release
year
2026
stack
Python 3.11, FastAPI, SQLite, RouteLLM, Anthropic + OpenAI SDKs
size
~2,000 lines Python · 8 test modules

LLM Router

A local proxy that makes Claude Code model-agnostic: easy prompts go to a cheaper model, hard ones to Claude.

LLM Router: Claude Code points at a local proxy on port 9099 that pins active tool chains, scores each request with length, content and RouteLLM signals, and sends scores of 0.5 or more to Claude Sonnet and lower scores to a cheaper backend, sharing context through SQLite. Claude CodeANTHROPIC_BASE_URL proxy :9099pin active tool chainstranslate Anthropic ⇄ OpenAIincl. SSE streams score0.3·length + 0.3·content+ 0.4·RouteLLM Claude Sonnet≥ 0.5, or tool_use Cheaper modelNIM·Groq·Ollama… SQLite context24 h TTL, shared
fig. 3 — routing one requestno extra keys → passes everything to Anthropic

Problem

Every Claude Code request goes to the most expensive model, even "rename this variable".

Approach

  • Composite score from length, content type and a RouteLLM matrix-factorization classifier.
  • Translator converts Anthropic Messages to OpenAI Chat Completions and back.
  • Routing log + dashboard show where traffic went.

Decisions

  • Once a tool-call chain starts, stay on that backend so agent loops don't break across models.
  • Any tool_use or tool_result block goes to Claude.
  • Ships as a Claude Code plugin with slash commands.
year
2025
stack
Python, Streamlit, LlamaIndex, Pinecone, Hugging Face embeddings, OpenAI GPT-3.5 Turbo
size
~190 lines Python · single-file app

CPS Parent Guide

A retrieval-augmented chat assistant that answers Chicago Public Schools parents' questions about enrollment, programs and policies, grounded in CPS material.

CPS Parent Guide: a parent asks a question in a Streamlit chat app; LlamaIndex embeds it with Snowflake arctic-embed, retrieves the top 10 chunks from Pinecone, packs them into a CPS-specific prompt with compact synthesis, and GPT-3.5 Turbo at temperature 0.3 writes the answer, falling back to the school or cps.edu when the context has no answer. LlamaIndex · RetrieverQueryEngine Parentplain-English question ask.py · Streamlitmulti-chat threadshistory per sessionuser's own OpenAI key Embed queryarctic-embed-m-v1.5 Pineconetop-10 CPS chunks Compact synthesisCPS parent prompt GPT-3.5 Turbotemperature 0.3 No answer in context?point to the school or cps.edu, don't guess answer
fig. 4 — answering one parent questionembeddings run locally; only generation calls OpenAI

Problem

CPS information on enrollment, transportation, meals and health forms is spread across many pages. Parents want a direct answer, not a site search.

Approach

  • CPS material indexed in Pinecone; each question pulls the 10 closest chunks.
  • LlamaIndex compact mode packs the chunks into as few model calls as possible.
  • Sidebar keeps separate chat threads, each with its own history.

Decisions

  • Open Hugging Face embeddings run in-process, so retrieval costs nothing per query.
  • Low temperature and a grounded prompt: when context is missing, send parents to the school, not a guess.
  • Users bring their own OpenAI key; it stays in the session and is never stored.
10chunks retrieved per question
10parent topic areas covered
2deploy paths: Streamlit Cloud, Heroku

Transformer encoder, from scratch

2026 · NumPy · 54 testsgithub ↗

A full pre-norm Transformer encoder with no framework. Each module is annotated with the CUDA kernel it stands in for, and a minimal reverse-mode autograd sits underneath.

The outputs are checked against nn.TransformerEncoder with copied weights, to floating-point tolerance. It models GPU kernels in NumPy; it doesn't run on a GPU.

kernels.py tiled matmul 32×32 tiles, shared-memory GEMM online softmax streaming max + sum (FlashAttention trick) fused layernorm · fused bias+GELU · fused residual+LN

Relevant Priors

2026 · Python, FastAPI, Dockergithub ↗

Given a patient's current imaging study, predict which prior studies a radiologist should compare. Only 23.8% of 27,614 pairs are relevant, so answering "no" already scores 76.2%.

I mapped descriptions to 35+ body regions with adjacency, then iterated on the errors. Sub-millisecond, rule-based.

accuracy %, 50 → 10026,326 / 27,614 correct

AI engineering

Agents, evaluation, model serving and MCP tooling. Code for these isn't public yet.

focus
LLM evaluation, guardrails, red teaming, responsible AI
runs in
CI

LLM eval and guardrails harness

Tests a RAG app for hallucination, prompt injection and PII leakage, with scores tracked in CI.

Eval suites for hallucination, prompt injection and PII leakage run against a RAG app; the outputs are scored and the scores are tracked in CI. eval suiteshallucinationprompt injectionPII leakage RAG appsystem under test scoringper test category CIscores tracked
fig. 6 — every change runs the suitesthree failure modes, one score history

Problem

A RAG app can look fine in a demo and still make things up, follow injected instructions or leak personal data. Without tests, regressions ship quietly.

Approach

  • Separate suites for hallucination, prompt injection and PII leakage.
  • Score the app's answers per category.
  • Track the scores in CI so a regression shows up on the change that caused it.
focus
multi-agent systems, Go, LLMOps, inference optimization
language
Go

Multi-agent research assistant in Go

A planner agent and worker agents with tool calling, plus a dashboard for cost and latency.

A research question goes to a planner agent, which hands work to several worker agents; workers call tools, and a dashboard tracks cost and latency. questionresearch task planner agenthands out work worker agenttool calling worker agenttool calling worker agenttool calling toolscalled by workers dashboardcost · latency
fig. 9 — plan, fan out, measurewritten in Go

Problem

Multi-agent systems get slow and expensive fast, and it's hard to see where the time and tokens go.

Approach

  • A planner agent splits the work across worker agents.
  • Workers use tool calling.
  • A dashboard tracks cost and latency.

Systems from my resume

No public repos for these, so no links. Details are as written on my resume.

Distributed key-value store

C++ · stream sockets

Led backend development of a fault-tolerant store behind cloud storage and webmail services, with dynamic tablet partitioning across storage nodes.

Primary-secondary replication, write-ahead logging and checkpoint recovery; LRU eviction under memory pressure.

NL-to-SQL with fine-tuned LLMs

PyTorch · Hugging Face · DeepSpeed

Fine-tuned LLaMA-7B and DeepSeek-Coder on 120,000+ natural-language-to-SQL pairs across multi-GPU A100 clusters; 30% better SQL correctness than the base models.

Domain-adaptable training for finance, e-commerce and healthcare dialects, plus a query frontend.

Data integrity verification

Cassandra · Ethereum · MapReduce

Tamper detection combining Merkle trees, Cassandra storage and Ethereum-anchored hashes; detected 94% of tampering across 1,200+ records.

Verification parallelized with MapReduce, 4× faster than the baseline.

More projects

Hover a row for the spec. Everything links to its repo unless it's private.

~/experience $ git log --oneline

Experience

From object detection on satellite imagery to payment-grade microservices to parsing database wire protocols. Open a role for the details.

    • Develop the Guardium Sniffer, a C++20 engine that captures and decodes live Oracle and MySQL network traffic for real-time database activity monitoring across enterprise deployments.
    • Extend ANTLR3 and BNF-defined SQL grammars to parse and normalize queries across dialects; use tcpdump and TCP/IP packet analysis to diagnose wire-protocol parsing failures.
    • Build containerized Podman test environments against live Oracle and MySQL, and automate regression validation with Python and Bash.

    C++20 · ANTLR3 · TCP/IP · tcpdump · Podman · Python · Bash

    • Engineered Java/Kotlin Spring Boot microservices and 20+ REST APIs handling over 10 million daily transactions; saga recovery and Redis idempotency kept correctness at 99.95% under concurrent load.
    • Optimized SQL/JPA queries across 300M+ monthly records, cutting peak database latency by 20%.
    • Built Kafka event pipelines with idempotent consumers across 12+ integrations: throughput up 35%, no duplicate-charge incidents. Tuned Prometheus and Istio to fix gRPC stream exhaustion: throughput up 37%, P95 latency down 12%.
    • Built and deployed a production RAG assistant using LoRA-fine-tuned LLMs on AWS GPU infrastructure with Terraform and FastAPI, serving 500+ engineers and 2,000+ daily queries; hallucinations down 41%, ticket resolution time down 33%.

    Java · Kotlin · Spring Boot · Kafka · Redis · gRPC · Istio · Prometheus · AWS · Terraform · FastAPI

    • Developed Python test suites for Coursera's autograder, reducing grading workload by 30%.
    • Guided 140+ graduate students through Python, SQL and machine learning problems, contributing to a 15% increase in student satisfaction.

    Python · SQL · machine learning · autograding

    • Built and optimized distributed PySpark ETL pipelines on Azure processing millions of records daily, cutting processing time by 25%; added data-quality checks and monitoring for downstream analytics teams.
    • Automated on-premises VM migrations to Oracle Cloud with the OCI Python SDK, reducing migration effort by 60%; extended a deployment tool with new OCI resource types and CMDB API integration.

    PySpark · Azure · Oracle Cloud (OCI) · Python

    • Designed a reusable, PyTorch-inspired deep-learning library with modular data-loading, training and inference; used it to improve object detection by 10% mAP across YOLO, RetinaNet and EfficientDet.
    • Built automated Python pipelines for satellite-imagery detection and LiDAR point clouds of 500M+ points, reducing manual inspection by 25% at 85% detection accuracy.

    PyTorch · YOLO · RetinaNet · EfficientDet · LiDAR

~/toolkit $ cat manifest

Toolkit

What I use, and where I've used it. No percentages.

languages

C++ (C++20)
Guardium Sniffer at IBM; distributed KV store
Java · Kotlin
Spring Boot microservices at ChiEAC
Python
FastAPI services, ML pipelines, Sentinel engine, test automation
TypeScript
Forge, MCP servers, VS Code extensions, Next.js apps
SQL · Bash
Query tuning on 300M+ monthly records; regression harnesses

backend & distributed

Spring Boot · FastAPI · NestJS
Production APIs, Sentinel, Villa
Kafka · Redis
Idempotent consumers, saga recovery, kill switches
gRPC · REST
20+ APIs; fixed gRPC stream exhaustion
Replication · WAL
Primary-secondary store with checkpoint recovery
MCP
Sentinel; whatsapp-mcp bridge fix

data & storage

PostgreSQL · MongoDB
Sentinel, Villa (Prisma), ShopMaster
Oracle · MySQL
Wire-protocol decoding at IBM
Cassandra · Hadoop
3-node cluster, MapReduce jobs, integrity verification
Spark / PySpark
Azure ETL at EAIESB
SQLite
LLM Router context store, Pulse

cloud & infrastructure

AWS · Azure · OCI
GPU inference, ETL, VM migration
Docker · Podman
Test environments against live databases
Terraform · Istio
RAG assistant infra; service-mesh tuning
Prometheus
Throughput and P95 latency work
GitHub Actions
CI and release pipelines for Sentinel, Forge

ai / ml

PyTorch · TensorFlow
Object detection at L&T; LSTM/GRU text models
Hugging Face · DeepSpeed
LLaMA-7B / DeepSeek-Coder fine-tuning
LoRA · RAG
Production assistant for 500+ engineers
scikit-learn
Census, HarmonySpeak, Relevant Priors baselines
NumPy (no framework)
Transformer encoder + autograd

developer tools

VS Code extension API
Pulse, Butter Bear (published)
Ink · Bun
Forge terminal UI and single-file binaries
ANTLR · tcpdump
SQL grammars and packet debugging
Claude Code
LLM Router plugin, daily pair programmer

~/about $ whoami

About

I like the layer underneath: the engine that decides, the proxy that routes, the parser that makes sense of bytes on the wire.

I studied computer and communication engineering at Manipal, then computer and information sciences at Pitt, where I also taught Python, SQL and machine learning to graduate students. Since then I've built payment-grade services, a production RAG assistant, and now database monitoring at IBM.

My approach is to make the system explain itself. Sentinel keeps a journal that explains every trade in plain language; Relevant Priors went from 58% to 95% by reading its own errors, one iteration at a time. When something breaks, I'd rather have a log that says why than a dashboard that says something.

Outside work I build tools for myself and publish the good ones. My editor theme, Butter Bear, has been going since 2021 and comes with a bear that gets hungry if you stop coding.

education

University of PittsburghMS, Computer and Information Sciences · 2023–2025 · GPA 3.957

Manipal Academy of Higher EducationBS, Computer Science & Communication Engineering · 2019–2023

open source

47 pull requests to other projects, 23 merged.

Fixed the WhatsApp bridge in whatsapp-mcp after a protocol change (PR open).

Butter Bear on the VS Code Marketplace and Open VSX.

interests

Agent infrastructure and MCP. Database internals and wire protocols. Interpretable models. Quantitative finance.

~/contact $ open roles in backend, distributed systems and AI infrastructure

Contact

tanukusai.sai52@gmail.com
Write an email