AI/ML ENGINEER · SYSTEMS BUILDER ·

I build the
infrastructure
other people's AI runs on.

AI/ML engineer who works below the prompt — caches, schedulers, voice pipelines, and the plumbing that keeps agentic AI fast and dependable. Most recently wrapped a contract at CodeWisp (YC W26), designing the LLM orchestration architecture behind their production agent workflows.

View the work Get in touch
sushanth@portfolio — zsh
sushanth@portfolio:~$ █
509ms END-TO-END VOICE RTT · 78% MEMORY SAVED VS. FLAT KV CACHE · <50ms BARGE-IN INTERRUPT CUTOVER · 500+ DSA PROBLEMS SOLVED IN JAVA · 4 SELECTABLE ENCRYPTION LEVELS, ONE INBOX · MCTS + LANGGRAPH RESEARCH LOOP ·

$ cat expertise.log

Not another prompt-engineer.

I'm drawn to the layer underneath the chatbot — the part that decides whether an agent responds in 500ms or 5 seconds, whether a cache hit saves a re-computation or silently corrupts state, whether a voice agent can be interrupted mid-sentence without falling apart. Most of my projects are an excuse to build that layer myself: a radix-tree KV cache, a duplex voice pipeline, a tree-search research agent, a quantum-secured email client.

Most recently, I wrapped a contract at CodeWisp (YC W26), designing the LLM orchestration architecture behind their production agent workflows. Outside of contract work, I spend a non-trivial amount of free time arguing with myself about eviction policies.

  • AGENTIC SYSTEMS Multi-agent orchestration, LangGraph workflows, MCTS-driven research agents, cost-aware model routing.
  • LOW-LATENCY AI Async, event-driven voice/agent pipelines with sub-second round-trips and aggressive interruption handling.
  • HIGH-PERFORMANCE SYSTEMS Concurrent data structures, custom caching layers, distributed systems in Go and Python.
  • APPLIED SECURITY Post-quantum cryptography, QKD-inspired key management, tiered encryption design.
  • ML / DL RESEARCH Tree search, ablation studies, experiment tracking, structured evaluation.
  • OPEN SOURCE Contributor to Taipy, distributed load-testing tooling, and Kafka tooling — 500+ DSA problems solved in Java.

$ ls -la ./projects

Four systems, one obsession: make agentic AI not fall over.

01 / 04 PYTHON

Bhasha-Stream

Sub-second duplex voice AI for code-switched Indian languages.

A production-style voice agent built for how Indians actually talk — sliding between Hindi, Tamil and English mid-sentence. An event-driven pipeline streams audio through VAD, ASR, an LLM and TTS in parallel instead of waiting on each stage in turn, with a four-state interruption model that lets the agent stop talking and start listening again in under 50ms when a caller speaks over it.

~509msround-trip latency
<50msbarge-in cutover
4-stageasync pipeline
FastAPIWebSocketsvLLMFaster-WhisperSilero VADMeloTTSDockerCUDA
VIEW SOURCE →
02 / 04 PYTHON

EvoThinker

An agent that proposes, runs, and critiques its own research.

A self-directed research agent that searches a tree of hypotheses the way a scientist iterates on experiments — propose an idea, write and sandbox the code to test it, score the result, then decide whether to go deeper or backtrack. Five cooperating LangGraph agents (proposer, architect, evaluator, critic, reporter) sit on an MCTS search core, closing the loop from idea to a written report with no human in the middle. Inspired by Sakana AI's AI Scientist-v2.

MCTS+ progressive widening
5cooperating agents
sandboxedexecution
LangGraphMCTSVector MemorySandboxed ExecPython 3.11
VIEW SOURCE →
03 / 04 PY · TS

QuMail

Email with a quantum-secure dial — pick your own paranoia level.

A desktop email client that treats encryption as a sliding scale instead of a single switch: an information-theoretically secure one-time pad for the most sensitive messages, QKD-derived AES-256-GCM for everyday secure mail, post-quantum Kyber + Dilithium for messages that need to stay safe decades from now, and a plain fallback for compatibility. A simulated key manager speaks the real ETSI GS QKD 014 protocol, so the security layer can be swapped for actual QKD hardware later.

4selectable security levels
ETSIQKD 014-compliant key API
Kyber-768+ AES-256-GCM
FastAPIReactTypeScriptElectronliboqsOAuth2
VIEW SOURCE →
04 / 04 GO

RadixKV

Stop paying to re-think the same prompt prefix.

When hundreds of agent branches share the same system prompt, most LLM caches still re-process those identical tokens on every single call. RadixKV fixes that with a concurrent radix tree that lets thousands of branches share memory for common prefixes — lock-coupled for safe concurrent mutation — paired with an O(1) min-heap eviction policy that prunes dead conversation leaves while keeping shared trunks intact.

~78%memory footprint saved
O(1)tree-aware eviction
gRPCcontrol plane
GogRPCRadix TreeMin-Heap LRUProtobuf
VIEW SOURCE →

$ cat system-architecture.yaml

Research depth, agent fluency, and the engineering to ship it.

core — AI, ML & research

PROGRAMMING

Python SQL Go TypeScript JavaScript Bash

MACHINE LEARNING

PyTorch TensorFlow scikit-learn NLP Computer Vision XGBoost Feature Engineering Fine-Tuning MLflow

GENERATIVE AI

RAG Agentic AI Multi Agent Systems LangGraph LangChain DSPy MCP Tool Calling Context Engineering vLLM

breadth — production engineering

BACKEND ENGINEERING

FastAPI AsyncIO REST APIs gRPC WebSockets PostgreSQL Redis Kafka Celery System Design Distributed Systems Concurrency

DATA ENGINEERING

Apache Spark Apache Airflow Hadoop DuckDB Polars ETL Pipelines Streaming Vector Databases

AI INFRASTRUCTURE

Docker Kubernetes NVIDIA Triton TensorRT ONNX CI/CD AWS Google Cloud

analysis — data science & research

DATA SCIENCE

Statistics Mathematics A/B Testing Forecasting Time Series Model Evaluation Experimentation

AI RESEARCH

Transformers Embeddings Retrieval Benchmarking Ablation Studies Optimization

FRONTEND

ReactTypeScriptNext.jsElectronTailwind CSSState Management

$ cat experience.log

Where the work happened.

2026

AI/ML Engineer (Contract) — CodeWisp YC W26

  • Architected and documented a production-grade six-layer LLM orchestration framework for autonomous software engineering agents, spanning planning, retrieval, execution, validation, memory, and response generation.
  • Designed the Advisor–Executor architecture that paired a Claude Opus strategic planner with Sonnet/Haiku execution agents, significantly reducing inference cost while maintaining high-quality reasoning.
  • Built progressive context loading, semantic retrieval, conversation memory, and knowledge assembly pipelines that dynamically selected relevant project context before every generation.
  • Implemented semantic caching, complexity-aware model routing, and token-budget optimization to eliminate redundant LLM calls, reduce latency, and improve overall inference efficiency.
  • Developed a two-stage code generation pipeline with AST-based validation, automatic error detection, structured regeneration, and output verification before delivery.
  • Integrated specialized AI agents for architecture planning, code generation, debugging, validation, documentation, project analysis, and workflow orchestration into a unified execution framework.
  • Authored a ~1,000-line production system prompt and comprehensive architecture documentation explaining system design, execution flow, optimization strategies, and engineering decisions for founders and stakeholders.
2025

Data Engineer & Analyst Intern — Elevate Labs

  • Designed Python- and SQL-based ETL pipelines to ingest, clean, validate, transform, and prepare operational datasets for downstream analytics and business intelligence workflows.
  • Automated recurring data engineering workflows by building reusable Python scripts and optimized SQL transformations, reducing manual reporting effort and improving data consistency.
  • Performed large-scale data profiling, schema validation, missing-value detection, duplicate resolution, and anomaly identification to improve overall data quality and pipeline reliability.
  • Built optimized SQL queries and reusable transformation pipelines to generate analytics-ready datasets powering dashboards, KPI tracking, and operational reporting.
  • Worked with relational databases to model structured datasets, optimize query performance, and maintain data integrity across multiple reporting sources.
  • Collaborated with cross-functional teams to translate business requirements into scalable data pipelines, automated reporting systems, and actionable analytical insights.

$ ./contact.sh

Let's build something that
actually has to work in production.

Open to AI/SDE internships and engineering roles — especially at fast-moving, YC-backed teams.

Email me LinkedIn GitHub