Open to work

Ships fast

AI Agents

Quant ML

Hello World 👋

Shresth
Samyak

> _

AI AGENT SAFETY★INFORMATION-FLOW CONTROL★MCP★TOKEN COMPRESSION★AST CALL-GRAPHS★QUANT ML★REALIZED VOLATILITY★QLoRA★JAX★LANGGRAPH★

[ 01 — ABOUT ]

WHO AM I?

I turn AI research into systems that ship and hold up in production.

I care about the unglamorous parts: what an agent is actually allowed to touch, how much context a model really needs, and whether a backtest is quietly cheating. Most of my work lives in the gap between a clever idea and a system that holds up under real inputs.

Right now that means a provenance control plane for AI agents, deterministic context compression for coding agents, and a volatility-forecasting layer for a covered-call fund. When I find a bug deep in a framework, I ship the fix upstream.

100%

attack containment

69%

token reduction

232

tests shipped

PyPI packages

Download Resume

[ 02 — EXPERTISE ]

WHAT I DO

Building the future ✦

tessera/policy.py

Lines: 0ANALYZING

Flow Trace

Agent Safety

AI Agent Safety & Tooling

Information-flow control between agents and their tools over MCP. Taint tracking, capability attenuation, and constrained plan interpreters so untrusted data cannot reach dangerous tools.

AST call-graph compression

LLM Systems

LLM Systems & Context

Shadow MCP servers and deterministic AST call-graph compression. Retrieval that returns identical output for identical queries, with zero model calls at retrieval time.

Realized Vol Forecast

[HAR-RV] h=1 refit

[GARCH] σ² updated

[SIGNAL] IV−RV gap

Vol Signal

IV: 24.1%

RV: 18.7%

VARIANCE GAP

142 bps

Quant ML

Quantitative ML

Realized-volatility forecasting (HAR-RV, GARCH/GJR, LightGBM), variance-gap signals, and purged, embargoed walk-forward cross-validation that refuses look-ahead leakage.

Job Stream

⚙ Job: enter patient record

ORCHESTRATOR

Queue: 3 jobs

Idem: ok

Retries: 0

Automation

Agentic Automation

Multi-agent workflows on LangGraph with interrupt/checkpointer state, plus deterministic desktop automation with idempotent orchestration and failure-evidence collection.

Detection

{

}

~50ms/frame99.4%

Vision

Applied Computer Vision

Fine-tuned detectors with multi-object tracking, embedding re-identification, and finite-state ownership logic running in real time on consumer GPUs.

✓

COVERAGE98%

Open Source

Open Source & Numerics

Debugging dtype-promotion and numerical-stability issues deep in ML frameworks, and shipping the fix upstream with a maintainer's review.

[ 02 — PROJECTS ]

WHAT I BUILT

AgentsSYS_ID // 001 · PUBLISHED

tessera · flow controlMCP

UNTRUSTED

GATE

TOOL

SCAN untrusted data → send_email

Tessera

A provenance control plane for AI agents. Information-flow control over MCP: untrusted data cannot reach an exfiltration-capable or irreversible tool without passing a constrained declassifier or human approval. Provenance taint tracking over a trust lattice, tool blast-radius classification, macaroon-style attenuating capabilities, and a CaMeL-style plan interpreter.

100%

containment

~½

utility cost

232

tests

PythonMCPCapabilitiesPyPI

GitHub PyPI

AI/MLSYS_ID // 002 · PUBLISHED

llm diet · ast compress946 nodes

0%tokens saved

32,041 → 32,041 chars946 kept

LLM Diet

Token compression for AI coding agents. A shadow MCP server intercepts Claude Code file reads and returns AST call-graph-compressed context instead of raw files. Deterministic: zero LLM calls at retrieval, identical output for identical queries. Benchmarked on a 946-node FastAPI project. Works with Claude Code, Cursor, and Windsurf.

69%

fewer tokens

32K→10K

chars/read

946

node bench

PythonASTMCPPyPI

GitHub PyPI

AI/MLSYS_ID // 003 · ACTIVE

self-correct · gen→critic→refineiter 1/3

> draft v1: ungrounded claim

MiniLM critic score0.38

verdict: REJECT

Self-Correcting LLM

Mistral-7B fine-tuned via QLoRA on a synthetically corrupted dataset. A Generator-Critic-Refiner loop uses a MiniLM critic for quality-score regression and verdict classification, plus browser-augmented grounding for hallucination detection.

QLoRA

3-stage

gen·critic·refine

Mistral-7BQLoRAMiniLMPEFT

GitHub

VisionSYS_ID // 004 · ACTIVE

surveillance · BoT-SORT~50ms

BAG#7

P12

owner dist: 0.0mscore 0.00

Surveillance

Real-time abandoned-bag detection on railway CCTV. YOLO fine-tuned on a custom dataset, BoT-SORT with BagRegistry embedding similarity for re-identification, FSM-based ownership tracking, and multi-factor abandonment scoring, all in real time on an RTX 4060.

~50ms

per frame

RTX 4060

consumer GPU

YOLOBoT-SORTOpenCVFSM

GitHub

QuantSYS_ID // 005 · PAPER

alphagrid · 8 agentsRISK: OK

P&L (paper)

+$1,240

Sharpe 0.9

funding · momentum · pairs

AlphaGrid

An 8-agent autonomous trading system across Indian equities (Angel One) and crypto (Binance). Every trade routes through a shared risk manager (hard caps, kill-switches, half-Kelly) and an append-only, tamper-evident track record. FinBERT news sentiment, cointegration pairs, a real-time WebSocket tick layer, and 196 tests.

agents

196

tests

½-Kelly

risk gate

PythonFinBERTRedisNext.js

GitHub Live

AgentsSYS_ID // 006 · PRE-ALPHA

agentpass · policy + audit<50ms

payments.send $4500ALLOW

database.deleteDENY

payments.send $1200ESCALATE

Ed25519 ⛓#a1–#b2–#c3–#d4

AgentPass

Compliance-ready identity, authorization, and audit for AI agents. A cryptographic passport (did:key + W3C Verifiable Credentials) per agent, a sub-50ms policy engine (allow / deny / escalate), and a hash-chained, tamper-evident audit log (Ed25519 + SHA-256). Adapters for LangChain, CrewAI, and the OpenAI Agents SDK, plus RBI/SEBI report templates.

<50ms

policy eval

Ed25519

signed log

adapters

Pythondid:keyPolicyAudit

GitHub PyPI

SystemsSYS_ID // 007 · PROD

coolroute · shade optimvalhalla

COOL 84% shade · +2.3minFAST 0% · fastest

CoolRoute

A heat-aware pedestrian routing engine that finds the shadiest path, not just the fastest. Real-time 3D shadow projection from solar azimuth/elevation, ML building-height prediction (XGBoost) for missing OSM data, and Valhalla routing scored against shadow polygons. React + MapLibre frontend, PostGIS, a multiprocessed ETL pipeline.

shadow engine

XGBoost

height ML

Valhalla

routing

PythonPostGISMapLibreXGBoost

GitHub

AI/MLSYS_ID // 008 · SHIPPED

product agentPRE_PURCHASE

washer for a family of 4?

🤖 AT-WM-9KG · fits your 60cm slot

OpenRouter · Mistral + Qwen-VL + Llama · never fakes specs

Echo — Product Agent

A dual-mode product intelligence agent. Pre-purchase it acts as a consultative salesperson (room-image analysis, color matching, fit checks); post-purchase it becomes a support engineer (error codes, troubleshooting, maintenance). Powered by free OpenRouter models (Mistral 7B + Qwen2-VL + Llama 3.1), FastAPI + React, and a safety-first stance that never hallucinates specs.

2-mode

pre / post

free LLMs

vision

room analysis

FastAPIReactOpenRouterVision

GitHub

[ 03 — JOURNEY ]

MY JOURNEY

Machine Learning Researcher

ML ResearchMythix — Quant Covered-Call Fund

Jun 2026 - Present

Designing the ML forecasting layer of a systematic covered-call options fund. HAR-RV, GARCH/GJR and LightGBM realized-vol models drive option-selling via a variance-gap signal (IV² − RV²). Built a feature pipeline from the implied-vol surface, VIX term structure, and multi-horizon realized vol, validated with purged, embargoed walk-forward CV. Audited the methodology and fixed a GARCH derivation error and a missing Itô drift term in the Monte Carlo simulator.

PythonHAR-RVGARCH/GJRLightGBMWalk-forward CV

AI Automation Engineer

AutomationGetHelpDesk.ai

Apr 2026 - Present

Building an automation layer that executes the full data-entry workflow inside dental PMS software (Open Dental, Dentrix) from structured patient and insurance data. Engineered a coverage-mapping engine translating abstract insurance tiers into ADA code-range configs feeding a deterministic pywinauto layer, with a state-machine JobOrchestrator handling idempotency, MySQL duplicate checks, and failure-evidence collection.

PythonpywinautoMySQLState machinesIdempotency

AI/ML Engineer — Freelance

FreelanceIndependent · Direct clients, Upwork, Fiverr

Aug 2024 - Present

Built a multi-agent procurement automation system on LangGraph (Sourcing, Risk & Compliance, Human-in-the-Loop Pricing) using interrupt/checkpointer state so buyer approval suspends and resumes graph state. Advising a US client building an agentic ticket marketplace on architecture and pricing for a wholesale arbitrage engine and a tactical short-selling sleeve.

LangGraphMulti-agentArchitecturePricing

Merged upstream

jax-ml/jax #37457

Numerics fix in dot_product_attention

Diagnosed a dtype-promotion overflow in masked attention: the float64 logit sentinel cast to −inf on the pre-softmax float32 conversion, tripping debug_infs before softmax normalized it. Capped at float32 max, preserving masking. Reviewed and approved by a JAX core maintainer.

Winner — Feature Creep Hackathon

Nov 2025

OCI Generative AI Professional

Oracle · 2025

OCI AI Foundations Associate

Oracle · 2025

[ 04 — CONTACT ]

LET'S TALK

Open to internships and research. Drop a note or reach me directly. I read every one. 👋

Direct lines

shreshthsamyak@gmail.com

+91 6299912442

Patiala, India

github.com/ShresthSamyak

linkedin.com/in/shresth-samyak

x.com/ShresthSamyak