Saurabh Singh

AskObserve

Production LLM · Conversational Analytics

accuracy

+35%

latency

−20%

search

<100ms

Spring-based microservice that puts an LLM in the loop for 25+ business units, with an intelligent selector that routes between Snowflake structured analytics and LLM/NLP sources.

• Built dual-source ML router that picks between structured analytics and LLM/NLP based on query shape
• Engineered sub-100ms semantic history search on Elasticsearch + MongoDB Atlas with compound indexes and fuzzy matching
• Designed tag-based filter pipeline (hierarchical teams, date-range merging, dynamic enrichments) translating UI selections into ML format
• AOP-based monitoring + cache-driven filter pipeline for low-latency reads

Hermes

Workflow Orchestration · Platform

events/day

20M+

tenants

1000+

code reduction

85%

storage saved

99.6%

Mission-critical platform transformation consolidating 15+ fragmented workflow implementations into a single Temporal-based orchestration layer.

• Designed a hybrid stage-group DAG with constrained inputs, enabling dynamic pipeline construction from compile-time constants
• Reduced new-stage addition effort from ~1 week to <1 day
• Eliminated 99.6% of Temporal storage overhead (~1TB → ~4GB/day) using a claim-check pattern
• Killed dual SQS/Temporal code paths — single source of truth for orchestration

Looker zero-downtime migration

Infrastructure · Cost

saved / yr

$144K

envs migrated

3 prod

downtime

zero

Architected and executed a zero-downtime Looker cluster migration across 3 production environments with a custom account-based routing framework.

• Custom account-routing framework for transparent cluster cutover
• Robust Dry-Run / Execute / Rollback APIs with state transitions in MongoDB
• Resolved auth-token cache collisions and large-dashboard payload limits via a custom Strip-Import-Patch approach

Vajra

Air-gapped Foundation Model Stack

block rate

100%

false positive

0%

tests

217

A solo-built sovereign LLM stack that runs fully offline on a laptop. Designed for high-stakes environments with classification-aware retrieval and tamper-evident audit.

• FastAPI gateway · loopback-only · cold-start <8s
• Foundation-model inference via llama.cpp (model-agnostic by design)
• Classification-aware RAG with metadata pre-filter applied before kNN — full role isolation across clearance tiers
• Layered input policy engine (regex / heuristics / optional offline classifier) with distinct reason codes
• Output validator: secret-pattern, PII, grounding-similarity, and system-prompt echo detection
• Tamper-evident audit log: append-only JSONL with SHA-256 hash chaining + standalone CLI verifier
• Frozen red-team corpus (152 attacks · 8 categories) with versioned scorecard

Cashfree Payments

Payments · Backend

test coverage

80%+

core svc PoCs

2

Backend features for critical onboarding flows. Led PoC migrations for core services (onboardingsvc, commonmerchantsvc) including Kong Gateway, ArgoCD, Kafka rotation, and DB migration.

• Rest Assured automation suite with 80%+ consistent coverage on service contracts
• Automated testing infrastructure on BrowserStack and AWS Grid
• REST API discipline on payments-grade onboarding flows