Principal Product Manager | Multi-Agent AI Systems | Enterprise Platforms | 15+ Years Building 0→1 at Scale

I'm a product leader who believes the best PMs are technical enough to build, strategic enough to influence executives, and hands-on enough to ship. I design and build multi-agent AI systems, enterprise data platforms, and real-time infrastructure products that deliver measurable business impact.

Currently driving AI product strategy at Microsoft (Teams Platform). Previously built enterprise data ecosystems at Starbucks serving 200+ data scientists globally.

Dual-agent system for product managers — structurally enforced ideation + adversarial review

• Architecture: Accomplice (divergent ideation) and Gauntlet (adversarial review) with a hard structural boundary — neither agent can perform the other's role, enforced architecturally not by convention
• Gauntlet: Fans out across 12+ adversarial review lenses — governance, compliance, data/telemetry, responsible AI, reliability, engineering feasibility, GTM, competitive — each returning structured findings with severity and go/hold/no-go verdict
• PM Memory Layer: Federates from 6 upstream source types — meeting intelligence, signal clustering, org knowledge, incident history, telemetry, service catalog — with provenance tagging (verified vs. inferred) and quality grading before reaching any agent
• Offline-capable: Pluggable model backend; single-shot and conversational ideation modes; Hardener stage synthesizes cross-lens conflicts into backlog-ready specs
• Tech Stack: Python, Claude API, multi-agent orchestration, federated memory architecture

# What it does in one line:
PM idea → Accomplice ideates → Gauntlet stress-tests across 12+ lenses → Backlog-ready spec

Two production AI agents covering the full incident lifecycle at Microsoft

• ICM Classifier: Defined inference architecture and prompt engineering framework using Claude API; owned 8-metric evaluation framework (classification confidence, false-positive rate, false-negative rate, escalation trigger accuracy, latency P50/P95, human review rate) before any production traffic
• Impact: Reduced incident triage from 5–7 days to 2 minutes across 100–150 monthly incidents — org-wide production deployment, not a pilot
• Model Selection: Claude over GPT-4 for classification reliability and audit trail traceability
• Townhall Support Agent: Gemini Flash — agentic orchestration pattern: continuous telemetry ingestion → quality signal detection → threshold evaluation → automated incident report generation during live events at 20K+ concurrent participants
• Combined system: Classifier handles async triage, monitoring agent handles live event ops — full incident lifecycle covered
• Tech Stack: Python, Claude API, Gemini API, REST API integration, pandas, openpyxl, python-pptx

# What it does:
Live event telemetry → Quality signal detection → Auto-drafted incident report before customer escalation
100–150 monthly incidents → AI classification → 2-minute triage (was 5–7 days)

Parametric cost estimator for residential general contractors — deterministic, auditable, human-in-the-loop

• Problem: Low-to-mid volume residential GCs need rough-order-of-magnitude pricing before drawings exist — existing AI takeoff tools (Togal, Kreo, Handoff AI) require drawings that don't exist at this stage
• Approach: Parametric/conceptual estimating — matches new projects against comparable historical SOVs using weighted similarity scoring across project inputs (sqft, bed/bath, garage, levels, foundation type)
• Cost buildup: Per CSI division (Concrete, Framing, Roofing, Windows & Doors, Finishes, Casework, Plumbing, HVAC, Electrical, Site/Earthwork) normalized as % of total cost — survives small historical samples better than flat $/sqft
• Output: Range, not point estimate — total cost range, derived $/sqft, division-by-division breakdown, each tagged with confidence level (high/medium/low) based on input correlation
• Design principle: Deterministic, not generative — core matching and cost math is rules-based arithmetic, fully auditable; contractor's proprietary cost data stays in their environment
• Tech Stack: Python, HTML interactive prototype, parametric matching algorithms

# What it does:
Project inputs (sqft, beds, garage, levels) → Comparable project matching → 
CSI division cost breakdown with confidence flags → ROM range estimate

Why this matters: Built for a real contractor solving a real pre-drawings pricing problem. Shows product thinking: range over false precision, confidence shown per division, human-in-the-loop by design — not a black-box number generator.

Semantic search using Gemini 2.0 Flash + RAG + FAISS vector indexing

• Goal: Understand how modern AI recommendation products work from the inside
• Tech Stack: Google Gemini 2.0 Flash, FAISS, RAG architecture, few-shot prompting, collaborative filtering, content-based filtering
• Product Thinking: Integrated OTT/affiliate monetization model into the architecture

"Suspenseful thrillers with strong female leads" → Vector search → Ranked recommendations

Battle-tested templates from shipping products at Microsoft & Starbucks

• PRD templates (technical, consumer, 0→1)
• Prioritization frameworks (Impact Score Model, Tech Debt vs Features)
• OKR definition and tracking frameworks
• Evaluation frameworks for AI product quality (accuracy, latency, confidence thresholds)
• Stakeholder templates (Business cases, RFCs, Launch checklists)

Current: Senior Product Manager @ Microsoft — Teams AI & Media Platform

• Authored FY2027 Media Convergence strategy aligning 9+ organizations into single architectural direction

• Shipped 1080p streaming for 20K+ concurrent participants (sub-2s P90 latency, protected $14M revenue)

• Built COGS and pricing framework: unlocked $14M opportunity, prevented $7.9M revenue loss

• Shipped two production AI agents: incident classifier (Claude API) + real-time monitoring agent (Gemini) Previous: Product Manager @ Starbucks — Enterprise Data Platforms

• Built 360 Data Suite from zero: governed data layer for Marketing, Operations, Finance globally

• Owned MLOps pipeline on Azure ML: 92% forecast accuracy in production

• Architected Identity Stitching framework using MDM principles across 7 source systems

• Drove 200+ data scientists to 75% daily active adoption through embedded delivery strategy Earlier: Technical Product Manager @ Lamb Weston — ERP & Supply Chain

• Owned order management workflows for global accounts (McDonald's, IHOP)

• Directed ERP system transitions for acquired companies globally

Languages:

AI / Agent Systems:

Data & MLOps:

Builder Tools:

• 💼 LinkedIn: linkedin.com/in/srikanthjaligam
• 📧 Email: srikanth.jaligam@gmail.com
• 🌐 Location: Seattle, WA (Open to remote)

Building multi-agent AI systems. Shipping things that run in production. Learning in public.