I work on the parts of software that are hard to get right the first time: retrieval pipelines that need to be correct, event loops that need to not drop connections, and agent systems that need to know when to stop talking. Based in India, currently focused on production-grade AI systems, distributed infrastructure, and the backend architecture underneath both.
Alongside systems engineering, I regularly practice Data Structures & Algorithms to strengthen problem-solving and software engineering fundamentals.
Interested in
| AI Systems | Agentic AI | Distributed Systems | Backend Engineering | Networking |
| Operating Systems | Retrieval-Augmented Generation | Data Structure and Algorithms | Databases | Developer Infrastructure |
| Current Focus | Production-grade retrieval and multi-agent reasoning for NyayaOS |
| Currently Learning | Consensus protocols, storage-engine internals, and large-scale agent orchestration |
| Building | A C++ in-memory data store from raw sockets up — NovaCache |
| Open to Collaborate On | AI infrastructure, RAG systems, and low-level networking/storage projects |
| Latest Interests | Cross-encoder reranking, knowledge graphs for legal reasoning, event-driven architectures |
FLAGSHIP PROJECT · ENTERPRISE-SCALE AI PLATFORM
An end-to-end AI system for legal reasoning: it retrieves authorities, verifies citations against source text, and runs a structured multi-agent debate before producing a drafted memorial — rather than a single-shot LLM answer.
%%{init: {'theme':'dark', 'themeVariables': {'primaryColor':'#161233','primaryTextColor':'#E5E7EB','primaryBorderColor':'#6366F1','lineColor':'#8B5CF6','secondaryColor':'#0f172a','tertiaryColor':'#0b0e1a','fontFamily':'Fira Code'}}}%%
flowchart LR
Q[Query] --> QU[Query Understanding]
QU --> SR[Semantic Retrieval<br/>Vector Index]
QU --> KG[Knowledge Graph<br/>Traversal]
SR --> AP[Authority Pool<br/>Construction]
KG --> AP
AP --> RR[Cross-Encoder<br/>Reranking]
RR --> DE{Multi-Agent<br/>Debate Engine}
DE --> PA[Prosecuting Agent]
DE --> DA[Defending Agent]
PA --> JA[Judge Agent]
DA --> JA
JA --> CV[Citation<br/>Verification]
CV --> MG[Memorial<br/>Generation]
MG --> OUT[Drafted Output]
Tech stack & key innovations
| Retrieval | Semantic vector search, hybrid dense + sparse retrieval, legal corpus indexing |
| Reasoning | Multi-agent debate engine with role-conditioned agents and a judge agent for adjudication |
| Grounding | Citation verification layer that checks generated claims against indexed source text |
| Knowledge | Domain knowledge graph for statute/precedent relationships, traversed alongside vector retrieval |
| Reranking | Cross-encoder reranking over the authority pool before it reaches the debate engine |
| Output | Structured memorial generation from verified, reranked evidence |
Key innovations
- Two-stage retrieval (semantic + graph) merged into a single authority pool before reranking, instead of relying on vector search alone
- A debate-structured reasoning loop, so conclusions come from adjudicated argument rather than a single forward pass
- Citation verification as a hard gate — claims that can't be traced to source text don't reach the final draft
AGENT ORCHESTRATION · LANGGRAPH
A trading research platform where specialized agents own specialized concerns — market state, news sentiment, risk, strategy — and a judge agent resolves their (often conflicting) outputs into a single execution decision.
%%{init: {'theme':'dark', 'themeVariables': {'primaryColor':'#161233','primaryTextColor':'#E5E7EB','primaryBorderColor':'#6366F1','lineColor':'#8B5CF6','secondaryColor':'#0f172a','tertiaryColor':'#0b0e1a','fontFamily':'Fira Code'}}}%%
flowchart TD
YF[Yahoo Finance<br/>Market Data] --> MA[Market Agent]
NW[News Sources] --> NA[News Agent]
MA --> SG[Signal Generation]
NA --> SG
SG --> RA[Risk Agent]
SG --> STA[Strategy Agent]
RA --> JA[Judge Agent<br/>LangGraph Orchestrator]
STA --> JA
JA --> EA[Execution Agent]
EA --> BT[Backtesting Engine]
EA --> PF[Portfolio Analytics]
Tech stack & key innovations
| Orchestration | LangGraph-based agent graph with explicit state transitions between agents |
| Agents | Market, News, Risk, Strategy, Judge, and Execution agents, each with a narrow, testable responsibility |
| Data | Yahoo Finance market data feed |
| Evaluation | Backtesting engine and portfolio analytics for measuring strategy performance over historical data |
Key innovations
- Risk and Strategy agents run independently and are only reconciled by a dedicated Judge agent — no single agent has unchecked authority over execution
- Signal generation is decoupled from decision-making, so new data sources can plug into the same agent graph
SYSTEMS PROGRAMMING · C++ · EPOLL
An in-memory key-value store built from the socket layer up: a non-blocking TCP server on an epoll event loop, a hand-written RESP parser, and a storage engine with TTL and persistence — no framework in between.
%%{init: {'theme':'dark', 'themeVariables': {'primaryColor':'#161233','primaryTextColor':'#E5E7EB','primaryBorderColor':'#6366F1','lineColor':'#8B5CF6','secondaryColor':'#0f172a','tertiaryColor':'#0b0e1a','fontFamily':'Fira Code'}}}%%
flowchart LR
C[Client] --> TCP[Non-blocking<br/>TCP Server]
TCP --> EL[epoll<br/>Event Loop]
EL --> RP[RESP Parser]
RP --> CD[Command Dispatcher]
CD --> SE[Storage Engine]
SE --> TTL[TTL Wheel]
SE --> PS[Persistence<br/>Snapshot / Log]
CD --> RE[Response Encoder]
RE --> C
Tech stack & key innovations
| Networking | Custom non-blocking TCP server, epoll-based event loop, no external networking library |
| Protocol | Hand-written RESP parser compatible with standard Redis clients |
| Storage | In-memory storage engine with TTL expiry and thread-safe access |
| Durability | Persistence layer for surviving restarts |
| Concurrency | Thread-safety across the storage layer under concurrent connections |
Key innovations
- Single-threaded event loop for the network layer keeps the RESP protocol handling free of lock contention, while the storage layer is still safe under concurrent access
- TTL expiry implemented without scanning the whole keyspace on every tick
|
REAL-TIME · ENCRYPTED A messaging platform with JWT-based authentication and end-to-end AES encryption, layered with AI-native features: summarization, translation, and sentiment analysis on top of real-time WebSocket messaging. Tech stack
|
BACKEND · MARKETPLACE A backend for a multi-vendor e-commerce marketplace — vendor accounts, catalog and inventory management, and order flows modeled as independent, composable services rather than a single monolithic store. Tech stack
|
|
AI Systems 
Backend 
|
Distributed Systems 
Databases 
|
Full development toolchain
%%{init: {'theme':'dark', 'themeVariables': {'primaryColor':'#161233','primaryTextColor':'#E5E7EB','primaryBorderColor':'#6366F1','lineColor':'#8B5CF6','cScale0':'#161233','cScale1':'#1e1b3a','cScale2':'#0f172a','fontFamily':'Fira Code'}}}%%
timeline
title Engineering Journey
2024 : Backend Foundations
: Django
: Spring Boot
: System Design
2025 : Distributed Systems & AI
: NovaCache
: Networking
: Agentic AI
: RAG
: NyayaOS
: AstraFund
2026 : Production & Infra
: Production AI Platforms
: Infrastructure
: Open Source
Trophy case
- Starred
shreekanthashokg-lang/Price-Predictor-ML-Pipeline— Jul 31, 2026 - Pushed to
sathwiksgjois/AuraChat---Advanced-AI-powered-Chat-System— Jul 31, 2026 - Pushed to
sathwiksgjois/AuraChat---Advanced-AI-powered-Chat-System— Jul 31, 2026 - Pushed to
sathwiksgjois/AuraChat---Advanced-AI-powered-Chat-System— Jul 31, 2026 - Pushed to
sathwiksgjois/astrafund— Jul 31, 2026
