SEP 14/AI SYSTEMS
HOW AGENTS ACTUALLY USE TOOLS
From model output tokens to sandboxed execution environments: deterministic JSON schema validation, error recovery loops, and safety rails.
6 MIN READ
I MAKE THINGSTO SEE WHATHAPPENS.
I MAKE THINGS
TO SEE WHAT
HAPPENS.
I build software, explore agent architectures, and try to understand what breaks when systems scale.
ENGINEERING / SYSTEMS / ML / AGENTS
~/workbenchmain
type
help or press /++++
it worked on my machine
curiosity
max→
status
stable*
*probably
confidence
100%*
*give or take
DO NOT PRESS
01/Selected Work
THINGS I'VE
BUILT.
01/FEATURED PROJECT
MULTI-AGENT
RESEARCH SYSTEM
A multi-agent research system that breaks complex questions into smaller tasks, searches for evidence in parallel, evaluates findings, and produces a structured final report.
ROLE
AI / BACKEND / FRONTEND
STACK
PYTHON / LANGGRAPH / FASTAPI / NEXT.JS
DESIGN DECISION
Used explicit graph-based orchestration so planning, research and evaluation remain observable and independently controllable.
EXECUTION LIFECYCLESTEP 00 / 04
00 /INTAKE
Receive multi-faceted user research inquiry and establish search constraints.
02/FEATURED PROJECT
PRODUCTION
RAG SYSTEM
A hybrid retrieval engine combining sparse keyword matching and dense vector search with cross-encoder reranking for low-latency, verifiable factual recall.
ROLE
AI / DISTRIBUTED SEARCH
STACK
PYTHON / FASTAPI / QDRANT / BM25 / LLAMAINDEX
DESIGN DECISION
Paired sparse lexical search with dense semantic embeddings to prevent hallucinations on exact identifiers while retaining natural language recall.
03/FEATURED PROJECT
AGENTIC SECURITY
ANALYST
An event-driven triage assistant that intercepts security telemetry, correlates threat indicators against live intelligence feeds, and orchestrates incident mitigation.
ROLE
SYSTEMS / SECURITY / LLMs
STACK
TYPESCRIPT / NODE.JS / LANGCHAIN / SPLUNK API
DESIGN DECISION
Implemented deterministic rule verification before granting LLM agents permission to invoke quarantine and policy mutations.
MORE WORK
ARCHIVE / 2024—202602/Lab
EXPERIMENTS & PROTOTYPES
Exploratory codebases, active research spikes, and technical curiosity.
001Browser-control agent
active
002C++ memory experiments
studying
003Retrieval benchmark
active
004Tiny Redis clone
planned
++++
EXPERIMENT_001SEP 2026
Browser-control agent
AI / SYSTEMS
Question:
How do we constrain autonomous browser tool-calls to prevent runaway DOM state?
Testing:
- ›Chrome DevTools Protocol (CDP) session lifecycle
- ›Dynamic accessibility tree pruning
- ›Deterministic action verification loops
What I'm testing:
Constrained browser automation agent running in sandboxed headless Chromium with token-budgeted DOM snapshots.
Key Verification:
- ›Chrome DevTools Protocol (CDP) session lifecycle
- ›Dynamic accessibility tree pruning
- ›Deterministic action verification loops
03/About
THE PERSON
BEHIND THE
WORKBENCH.
TRAJECTORYTAP / HOVER NODE
→
→
→
→
→
→
→
AGENTS›tool use · memory · planning · evaluation
Most of the things I learn follow the same pattern: I find something I don't understand, pull at the thread, and eventually try to build something with it.
THE STACK CHANGES.
THE CURIOSITY DOESN'T.
Right now those threads happen to be agents, security, systems and lower-level software. A few months from now, there will probably be another one.
04/Capabilities
TOOLS I
REACH FOR.
Evidence-linked technical foundations. Hover any tool to inspect verified project or lab usage.
01
AI SYSTEMS
PRIMARY FOCUSAutonomous agent workflows, deterministic tool routing, and evaluation pipelines that keep LLM systems observable and reliable in production.
////
02
FULL-STACK WEB
COREFast, typography-driven web applications with high visual craft, clean component architectures, and minimal client-side overhead.
////
03
BACKEND / DATA
PRODUCTIONLow-latency retrieval engines, asynchronous worker queues, and structured data indexing pipelines built to handle edge cases gracefully.
////
04
SYSTEMS / SECURITY
EXPLORINGInvestigating runtime internals, memory layouts, network protocols, and automated threat triage pipelines.
///
05/Notes
THINGS WORTH
WRITING DOWN.
VIEW ALL WRITING↗blog.samyakgupta.com
AUG 28/SYSTEMS
UNDERSTANDING MEMORY IN C++ & CACHE LOCALITY
Why mechanical sympathy matters in modern software: cache lines, pointer indirection costs, and struct-of-arrays memory layouts.
8 MIN READ
JUL 19/RETRIEVAL
THE PRACTICAL LIMITS OF DENSE EMBEDDINGS
When vector search degrades on exact identifiers, alphanumeric codes, and domain queries, and why hybrid BM25 pipelines remain critical.
5 MIN READ
EXPLORATION ARCHIVE
READ FULL ESSAYS & BENCHMARKS↗
blog.samyakgupta.comLOOKING FOR THE
LESS INTERACTIVE
VERSION?
RESUME.PDF↓
1 PAGE / UPDATED SEP 2026
06/Contact
GOT SOMETHINGINTERESTINGTO BUILD?
GOT SOMETHING
INTERESTING
TO BUILD?
A role, project, architecture problem, or rabbit hole worth exploring together.
STATUS
OPEN TO OPPORTUNITIES
LOCATION & TIMEZONE
NEW DELHI, INDIA
UTC +5:30
DIRECT CHANNEL