Skip to main content
FRONTIER AGENTIC AI · 75+ SERVICE SIMS · ADK & MCP
SDE-3 · Frontier AI Lab Engagements (via Turing)2025 - Present

Agentic Tool-Use: 75+ Service Simulations, Ergonomic Toolsets & ADK Plugins

Across Frontier AI Lab engagements via Turing, Pulkit Mundra engineered stateful Python, Pydantic v2 (StrictBaseModel), and SQLModel/SQLite service simulations spanning 75+ enterprise SaaS surfaces. Consolidating sprawling V1 APIs into flat @tool_spec overlays (60 -> 18 tools) and summary-first projections cut schema tokens by ~50-82% and response payloads by ~78-83% at ~99% task solvability parity.

60->18Consolidated @tool_spec Overlays-80%Payloads & -50-82% Schema Tokens~99%Solvability Parity (0.0pp Loss)
System Architecture Flow
01Ergonomic @tool_spec Overlays & FD Stripping
0275+ Stateful Simulations (Pydantic v2 + SQLModel)
03Google ADK SkillToolset + Decimal Compute
04State-Based Multi-Turn Evals (Docker + Pytest)

Interactive Agent Tool and MCP Token Optimizer Playground

Interactive Benchmark · Clean-Room Public API Schema Study

Verbose V1 API Surfaces vs Ergonomic Toolset Overlays

Pattern:
TASK: “Search unread billing threads from this week, read the dispute email, and archive resolved notices.”
4-Turn Agent Context Footprint ComparisonThis pattern: -82% per 4-turn run (16,872 tok saved)
Email Toolset Overlay (Public Spec Study)
540 payload tok
// Ergonomic @tool_spec Overlay: 18 workflow-aligned tools
// Summary-first search (-83% per-call) + concise mutators (-88%):
search_emails({ query: "is:unread billing", max_results: 10 }) -> [
  {
    id: "1962a8f",
    sender: "billing@vendor.io",
    subject: "Invoice #8841 adjustment",
    snippet: "Updated credit memo attached..." // 87 tok/call
  }
]
get_email({ id: "1962a8f" }) -> { body_text: "...", attachments: [] }
trash_email({ id: "1962a90" }) -> { id: "1962a90", status: "trashed" }
Response Payload
540 tok
-81% net response tokens
Schema Footprint
642 tok
-83% static schema tax
Eval Outcome
~98.9% Parity
0.0pp loss (state-based eval)
Call-Count Tradeoff
4.1 calls (thin search + hydrate)
Pydantic v2 StrictBaseModel
Scale Cost Calculator (1M / mo)
~$42,180/mo(16,872M tokens saved)

Per-pattern numbers above are illustrative clean-room scenarios. Across 75+ production tool simulations the average was ~80% fewer response payload tokens (78-83% range) at ~99% solvability parity.

01 / Architecture & Systems

Engineering & Implementation

  • Engineered and maintained high-fidelity, stateful API simulations across 75+ enterprise SaaS surfaces (Gmail, Calendar, Slack, Jira, Shopify, Canva, GitHub, Cursor/Terminal) using Python, Pydantic v2 (StrictBaseModel), and SQLModel/SQLite.
  • Consolidated verbose V1 APIs into workflow-aligned Python @tool_spec overlays (e.g., 60 email endpoints -> 18 tools, 12 device-action tools -> 7) and Function Definition (FD) schema-stripping configs, cutting static schema tokens by ~50-82% and runtime response payloads by ~78-83% while maintaining ~99% task solvability parity (0.0pp accuracy degradation).
  • Built cross-service agent plugins on Google ADK (SkillToolset progressive disclosure) with MCP connectors, deterministic Python Decimal aggregation tools (eliminating LLM arithmetic hallucinations and cutting multi-turn prompt tokens by ~32%), and an interactive Gradio runner.
  • Developed containerized multi-turn coding benchmarks (Docker + Pytest), diagnosed legacy tool-name assertion bugs in refactored eval suites, and collaborated on complex open-source Python codebase reasoning, cutting per-service delivery from days to hours with reusable Cursor skills.
02 / Production Outcomes

Measurable Impact

  • ~78-83% smaller response payloads, ~50-82% fewer schema tokens, ~32% lower multi-turn plugin tokens, and ~99% solvability parity across 75+ service simulations
BENCHMARK & ARCHITECTURE MATRIX
Scroll ->

Production Agent Toolset Refactoring Outcomes Across Multi-Hop Evaluations

Production Agent Toolset Refactoring Outcomes Across Multi-Hop Evaluations
Architecture DimensionLegacy 1:1 V1 API Tool DumpErgonomic @tool_spec + ADK Overlay
Exposed Email/Calendar Tool Surface60 raw CRUD & admin endpoints18 flat workflow-aligned tools (-66% schema)
Search Response Payload Size~516 tokens/call (full MIME + headers)~87 tokens/call (-78% to -83% net trajectory)
Mutating Tool Echo (Send / Trash)~6.8 KB full resource snapshotThin status confirmation (-88% to -99%)
Cross-Service Financial AggregationRaw JSON paginated into LLM contextGoogle ADK SkillToolset + Python Decimal (-32%)
State-Based Benchmark SolvabilityBaseline reference~99% parity (0.0pp true accuracy loss)
03 / Technical Q&A

Frequently Asked Questions

01How do you test multi-turn AI agents deterministically without flaky live third-party APIs?+
By building stateful local digital-twin MCP servers backed by SQLModel/SQLite and strict Pydantic v2 input/output contracts (`extra='forbid'`), paired with containerized Docker + Pytest graders that assert against final database state rather than intermediate tool-call order.
02Why did consolidating 60 API endpoints into 18 MCP tools keep ~99% task solvability?+
34 rarely used admin/S/MIME endpoints only added attention noise, while merging 26 core operations into 18 flat, intent-aligned tools with clear @model_validator error messages eliminated parameter-guessing errors and saved ~66-82% of static schema tokens.
Open to Senior / Staff Roles & High-Impact Consulting

Building LLM agents or MCP servers that need higher accuracy and lower token spend?

I design production MCP & ADK toolkits, high-throughput Python/Django APIs, and reliable fintech & Linux software. Reach out directly and I typically reply within 4 hours.