Make the request legible
Turn ambiguity into goals, constraints, and measurable acceptance criteria.
DIRECT CHAT / PRIVATE CHANNEL
A quieter way to reach someone.
Start a private one-to-one conversation using the other person's exact User ID. No public directory. No friend requests.
SYSTEMS / PRODUCTION GENAI
Lokesh Reddy V.
I build AI systems that earn their way into production.
RAG, tool-using agents, evaluation infrastructure, and LLM security—designed for accuracy, auditability, and operational control.
LIVE TELEMETRY / REFERENCE SYSTEM
RELIABILITY99.2%
ACTIVE SYSTEMS04
LAST RELEASE2026.09
MODULES / TRUSTED DELIVERY
Can this be trusted in production?
Interrogate the evidence
Retrieve what matters, challenge assumptions, and expose missing context.
Choose the safest path
Compare quality, latency, cost, and operational risk before committing.
Ship with proof
Grounded, validated, traceable—and built with a recovery path.
ENGINEERING STUDIO / INTERACTIVE
Don't just read the portfolio.
Open the system, challenge the controls, replay the reasoning, and inspect the evidence behind the work.
Enter the Engineering Studio →TTY / 01
15 browser-only modules · no data upload
PRODUCTION CONSOLE / TELEMETRY
A release earns its confidence.
Hybrid retrieval92.4%Source coverage
Vector + lexical search with metadata filters, reranking, and lineage.
PRODUCTION READINESS / RELEASE CONSOLE
Evidence before rollout.
lokesh@genai / release-consoleILLUSTRATIVE SIGNALSPASSEvidence coverageCitations required
PASSTask evaluationRegression checked
PASSOutput contractSchema valid
WATCHWeak evidenceEscalate or refuse
Approved deployments only
Versioned prompts, bounded context, timeout, and cost controls.
Untrusted input stays untrusted
Inputs, retrieval, tools, and outputs are validated separately.
Quality travels with every release
Latency, security, evaluation, and regression checks stay attached to rollout.
WHAT I BUILD / CAPABILITY MAP
Reliable intelligence, not demo magic.
The engineering around a model determines whether it can be trusted in production.
Grounded knowledge systems
Hybrid retrieval, reranking, source lineage, and evidence-aware responses.
Bounded agent workflows
Explicit states, scoped tools, approval gates, retries, and verifiable outcomes.
Secure AI applications
Trust boundaries, injection defense, PII handling, output policy, and audit trails.
Evaluation & operations
Task suites, quality regression, latency, token cost, monitoring, and release gates.
SYSTEMS / REAL CONSTRAINTS
Built to be inspected.
01Azure OpenAI · LangGraph · Pinecone02Azure OpenAI · LangGraph · Pinecone03LangGraph · GPT-4o · LlamaIndex04RAGAS · DeepEval · Python
Regulated Financial Services Intelligence Platform
Surfaced grounded, cited answers in under 30 seconds versus 45+ minutes of manual search.
Professional Services Audit Intelligence Platform
Reduced manual document-review effort by 34%; expanded evaluation coverage to 87% and blocked 11 regressions.
Deal Advisory Due Diligence Agent
Reduced first-pass due-diligence summary turnaround from 5 business days to 1.5 days across 14 pilot engagements.
LLM Evaluation & Quality Gate Framework
Reduced post-deployment production regressions by 67% over a 12-month measurement window.
INDEPENDENT APPS / DESIGNED & BUILT
From the system
to the screen.
Native media across desktop, phone and TV. A WebKit browser with workspaces, profiles and a Rust core. Network control with your own server and visible evidence. Personal products with source and release details you can inspect.
01 / Native media player
StreamSetu
Available to downloadYour streams. One home.
Connect M3U, Xtream Codes, Jellyfin, Emby, Plex, and more. Browse your connections together without signing up for a StreamSetu account.
Mac · iPhone & iPad · Android · Fire TV & Android TVExplore
02 / Native web browser
BrowserSetu
Release candidateA little more room for the web.
Mac layouts, workspaces, nested groups, linked split scrolling and session checkpoints help organise and recover your browsing.
Mac · iPhone & iPadExplore
03 / Network control & privacy
VayuSetu
Mac 1.0 · iPhone submittedYour network. Your control.
Import WireGuard files, text or QR codes, configure WireGuard manually, or add IKEv2/IPsec. VayuSetu does not sell VPN service or run VPN servers.
Mac · iPhoneExplore
INTERACTIVE ARCHITECTURE / REQUEST PATH
Follow a request through the system.
03 / RETRIEVAL
Hybrid search with filters, query expansion, and lineage.
Control: Evidence lineage · Failure mode: Weak recall
ABOUT / ENGINEERING APPROACH
I work on the layers that make AI useful after the demo.
My path runs from transformer-based NLP and full-stack AI applications to retrieval systems, bounded agents, evaluation, and production operations. Today I build enterprise GenAI platforms where model capability is connected to workflow state, controlled tools, approval boundaries, and measurable release evidence.
Read my engineering philosophy ↗Build for evidence
Grounded answers, traceable decisions, and measurable quality.
Bound autonomy
Explicit states, scoped tools, approval gates, and recovery paths.
Operate responsibly
Security, privacy, observability, and release discipline from day one.
EXPERIENCE / ENGINEERING RECORD
From applied NLP to production GenAI.
01Sep 2026 — Present02Jan 2024 — Aug 202603Aug 2021 — Dec 202204Nov 2018 — Dec 2019
Generative AI Engineer
Tata Consultancy Services · United States
AI Engineer — AI/ML & Agentic Systems
Professional Services Organization · United States
Full Stack AI Engineer
BT Group · United Kingdom
Application Development Engineer — AI/ML Practice
Tech Mahindra · India
TECHNICAL SYSTEM / CAPABILITY MAP
Skills connected to evidence.
Models & Orchestration
GPT-4o · Azure OpenAI · Llama 3 · Claude 3.5 · LoRA / QLoRA · Prompt engineering · Tool calling · Structured outputs · LangGraph
Retrieval
Hybrid search · Cross-encoder reranking · Metadata filtering · Source attribution · Pinecone · FAISS · ChromaDB · Weaviate
Evaluation
RAGAS · DeepEval · 200+ curated test sets · Faithfulness · Answer relevancy · Regression suites · Red teaming
Platform
Python · FastAPI · REST APIs · SQL · MongoDB · Docker · Kubernetes HPA · Azure · Databricks · PySpark · MLflow
Security
Prompt injection defense · PII detection and redaction · Access-aware retrieval · Tool authorization · Output validation · Audit logging · Content controls
Operations
Latency and token metrics · Traceability · Version tracking · Monitoring · Feedback loops · Failure recovery
FIELD NOTES / ENGINEERING WRITING
Short arguments from the systems layer.
01Tool Calling · Agents02Fine-tuning · QLoRA03Security · Prompt Injection04Agents · Architecture
LLM Tool Calling: The Complete Field Reference
A decision-first reference for reliable tool definitions, parallelism, failure handling, provider adapters, argument validation, and multi-turn agent loops.
What Actually Happens When You Fine-Tune an LLM on 12,000 Domain Samples: A Production Account
A granular account of QLoRA on Llama 3 8B: the failed runs, the measured gains, the serving surprises, and why fine-tuning was not the system bottleneck we thought it was.
Prompt Injection: The Vulnerability Baked Into the Architecture of Every LLM Application You Will Ever Build
A structural investigation of mixed-trust context, indirect injection, partial defenses, privilege separation, and the residual risk that cannot be patched at the model layer.
Everything Around the Model: The Real Engineering of Production AI Agents
Why production agents live or die by planning, state, memory, tools, execution, and operational controls around the model.
OPEN CHANNEL / 2026
Need an engineer who thinks beyond the model call?
Open to roles and technical conversations around production GenAI systems.
Start a conversation ↗