Agent

How Mochi works

Mochi is the small assistant that lives inside a Karrou profile. Ask it a question, and it answers from the same /aryamanchhikara.md document an agent would read. Below is one real request replayed from a captured trace — the prompt as sent, the tokens as they came back, and the grounding line each sentence of the answer drew from.

Captured August 21, 2026 · model nvidia/nemotron-3-ultra-550b-a55b:free · question “What does this person do?”

1. What the model receives

A question to an agent is mostly document. The system prompt sets the rules; the profile markdown is dropped in whole as untrusted data; the question itself is the smallest piece. Tap a segment below to see the real contents.

1,112 prompt tokens reported by nvidia/nemotron-3-ultra-550b-a55b:free. Bar is proportional to characters — the token split varies per tokenizer.

2. How the answer arrives

Tokens stream back one chunk at a time. This replay uses the exact per-token timings the provider returned — nothing is faked. This endpoint did not return the alternatives it weighed, so hover shows only timing.

nvidia/nemotron-3-ultra-550b-a55b:free0/44

waiting for first token…

0ms / 6333ms

this endpoint did not return top_logprobs, so alternatives aren't available. Timing and token order are still real. (model: nvidia/nemotron-3-ultra-550b-a55b:free)

3. Where each sentence came from

Every sentence of the answer is matched back to the line of the profile markdown it draws from. Hover a sentence to see the source. Sentences with no match above a similarity threshold are marked as unattributable — the model claims something the .md doesn't directly support, and calling that out is more honest than pretending it's grounded.

answer

Aryaman Chhikara is a customer-facing technical practitioner and solo founder with 5+ years leading end-to-end enterprise engagements for 1M+ users across 20+ system integrations. He currently works as a Techno-Functional Business Analyst at NYSLRS, acting as a technical trusted advisor who owns the full lifecycle from discovery through production rollout, and builds onboarding playbooks, debug runbooks, and demo assets. He's also building Karrou, a creator-portfolio SaaS platform with OAuth-connected widgets, LLM-grounded retrieval, and a customizable profile canvas. His daily toolkit includes Claude Projects, Claude Code, MCP connectors, and full-stack development with Next.js, Supabase, and TypeScript.

hover shows sourceunattributable

aryamanchhikara.md

1> Canonical machine-readable profile for aryamanchhikara. Structured for parsing.
2 
3# Aryaman Chhikara
4 
5Username: aryamanchhikara
6Headline: Customer-facing technical practitioner and solo founder who treats AI fluency as a craft
7Type: professional
8 
9## About
10 
11Customer-facing technical practitioner and solo founder with 5+ years owning end-to-end enterprise customer engagements serving 1M+ users across 20+ system integrations. Daily Claude Projects user with persistent context workflows and documented review processes. Comfortable translating between executive and engineering contexts to produce reusable playbooks and demo assets.
12 
13Working on: Building Karrou, a creator-portfolio SaaS platform with OAuth-connected widget integrations, LLM-grounded retrieval over structured user data, and a customizable profile canvas with CSS-variable theming and animation presets
14 
15## Experience
16 
17### Techno-Functional Business Analyst — NY State and Local Retirement System (NYSLRS)
182023-07 to present
19Acts as technical trusted advisor across 15+ enterprise stakeholder relationships, owning full lifecycle from discovery through tailored solution design, POC validation, executive readout, and production rollout. Leads structured discovery sessions with business stakeholders and IT, translating ambiguous operational problems into clear technical narratives proven against real workflows and data. Architects and debugs cross-system integrations across 20+ interconnected production systems, including tracing a downstream data pipeline defect rooted in a date calculation error through the full data flow to a confirmed fix. Built the team's onboarding playbooks, configuration reference patterns, and reproducible debug runbooks from scratch, adopted as standard for ramping new engagements. Serves as technical voice of customer back to engineering, surfacing patterns from live engagements that shipped concrete fixes and capability changes. Manages live Zendesk ticketing workflow with 10+ parallel open threads, prioritizing by business impact and translating between executive and engineering language. Pulls production data with SQL to prepare customer-specific scenarios, surface adoption signals, and frame findings as solution narratives for senior stakeholders.
20 
21## Skills
22 
23- Claude Projects (persistent account context, multi-step
24- Claude Code (daily user, shipping in production)
25- MCP connectors
26- Anthropic Claude API
27- OpenAI API
28- OpenRouter
29- prompt engineering
30- RAG patterns
31- tool-use and function-calling design
32- deliberate review of AI output before external use
33- Technical discovery
34- tailored demo design and delivery
35- POC scoping against real customer data and success criteria
36- demo environment and sample dataset construction
37- executive-level technical readouts
38- enablement documentation and playbook authoring
39- technical objection handling
40- Camtasia for demo recording
41- REST APIs
42- JSON Schema
43- webhooks
44- Postman
45- data import/export patterns
46- SQL (Oracle, warehouse-style queries)
47- spreadsheets and structured data prep for POCs
48- cross-system data flows across 20+ interconnected
49- Python
50- TypeScript
51- JavaScript
52- React
53- Next.js 14 (App Router)
54- Node.js
55- Supabase
56- NextAuth
57- Tailwind
58- Vercel
59- Git
60 
61---
62 
63Source: https://karrouhq.com/aryamanchhikara
64

What this is, and isn't

This shows how Mochi selects and generates, from a real captured call. It does not show the model's internals — those live on the provider's hardware and no API exposes them. If you want to see activations, attention heads, and residual streams on a small local model, there's a separate space for that:

transformer internals on Hugging Face → (a smaller, different model — the point is to look inside one you can actually run locally.)

Regenerating the fixture is a deliberate act: it costs one call against Karrou's free-tier quota. See scripts/capture-mochi-trace.mjs.