CASE FILE · 002 · ERUDITE V2 · PENDING RESOURCES

The guidance counselor
Bangladeshi applicants
never had.

A bespoke admissions intelligence system that replaces sterile chatbot responses with forensic financial aid auditing, structured visual counsel, and unapologetic reality checks — built specifically for high-need students navigating elite US admissions from Dhaka, Sylhet, and Chittagong.

70+
PAGES PER UNIVERSITY REPORT
$16K
MEDIAN AID MISPRICING GAP
Need-Blind
SCHOOLS CORRECTLY MAPPED
SSC/HSC
CURRICULUM BARRIER DECODED
Forensic financial aid auditing Structured visual counsel SSC/HSC contextualization Need‑blind endowment mapping Adversarial self‑audit Cryptographic provenance chain
§ 01 — THE ASYMMETRY

Generic advice doesn't just fail Bangladeshi applicants.
It actively harms them.

Every year, thousands of high-achieving Bangladeshi students apply to US universities using advice calibrated for American suburban applicants. The infrastructure of guidance — school counselors, college prep courses, alumni networks — simply doesn't exist here in the same form. What fills the gap is generic AI advice and marketing material written for a different student entirely.

The consequences are predictable: financial mispricing, curriculum gate blindness, asymmetric risk ignorance, and wall-of-text fatigue. Erudite V2 exists because the standard playbook assumes infrastructure, context, and emotional bandwidth that Bangladeshi applicants do not have.

FINANCIAL MISPRICING

Students see "$82k average international aid" in marketing materials and budget accordingly — not realizing the actual institutional award median is $66k. A $16,000 gap per year changes enrollment decisions.

CURRICULUM GATE BLINDNESS

Universities say they "accept global curricula" but provide zero explicit guidance for SSC/HSC grading contexts. Applicants get filtered before holistic review — not because of merit, but because no one told them how to translate their record.

ASYMMETRIC RISK IGNORANCE

A student with a 1430 SAT applies to three low-endowment safety schools and one reach. The safer financial bet — a need-blind, full-need institution like Amherst or Williams — gets skipped because selectivity looks scarier than sticker price.

WALL-OF-TEXT FATIGUE

Existing AI counselors dump dense paragraphs on anxious 17-year-olds. The information may be technically correct but psychologically inaccessible at the moment of highest stress. The format is the failure.

"Not just accurate answers — structured visual counsel. Every response is parsed into scannable badges, interactive charts, and collapsible detail panels. The UI is the intervention."

§ 02 — DESIGN PHILOSOPHY

Not a chatbot.
A structured counsel engine.

✗ STANDARD AI COUNSELOR
Flat text responses
Generic encouragement ("You can do it!")
Marketing-headline financial data
One-size-fits-all curriculum advice
No verification layer
Session amnesia
Tone: cheerful assistant
✓ ERUDITE V2
Structured JSON → rendered as cards, charts, accordions
Unapologetic data-backed reality checks
Forensic CDS audit with actual award medians
Bangladesh-specific contextualization flags
Adversarial self-audit on every response
Cryptographic provenance chaining per session
Tone: compassionate older sibling who's been through it
§ 03 — DATA LAYER

Eliminating the
JSON schema apocalypse.

Traditional RAG systems force complex, irregular admissions datasets into rigid schemas. Every time a university alters a minor policy, the database breaks. Erudite V2 stores Deep Research Reports — 70+ pages per university — in raw markdown format directly in Cloud Storage. No ETL. No schema migration. No fragility.

Discovery Engine's serverless layout parser ingests, chunks, and indexes raw files via native parsing (data_schema="content"). Semantic search retrieves exact paragraphs, metrics, and CDS citations without a managed database. Zero Spanner costs. The nuance that structured schemas destroy is preserved entirely.

RAW MARKDOWN STORAGE
70+ page Deep Research Reports per university stored without parsing or ETL. Policy changes require a file update, not a schema migration.
DISCOVERY ENGINE NATIVE PARSING
Semantic search retrieves exact paragraphs, metrics, and CDS citations without a managed database.
ZERO IDLE COST
Serverless storage natively. No Spanner. No idle hourly billing. The $1,000 Gen AI App Builder credit absorbs Discovery Engine retrieval.
CREDIT-SAFE RETRIEVAL
Standard Discovery Engine Search API performs state-of-the-art semantic retrieval without calling Vertex AI prediction meters.
§ 04 — THE COUNSELOR YOU NEVER HAD

Three linguistic constraints.
Enforced at the prompt level.

CONSTRAINT 01 · TONE

Calm, selective, authoritative, deeply realistic. Never uses AI filler — no "Sure, I can help with that!", no "Great question!". Speaks with quiet, unapologetic data-backed confidence. The voice of someone who has read the actual CDS, not the admissions brochure.

CONSTRAINT 02 · FINANCIAL VERIFICATION

If a student cites marketing-headline financial data, the system corrects it against audited CDS figures before proceeding. Budget around the mid-60s, not the $82k headline. A 4-year guarantee is worth more than a slightly larger freshman year grant.

CONSTRAINT 03 · REALITY CHECK PROTOCOL

If a student with a weak test score gravitates toward low-endowment safety schools, the system gently but firmly deconstructs the asymmetric risk. Full-need endowments are actually safer financial bets despite extreme selectivity. Over-explain your SSC/HSC grading context or risk being filtered before holistic review begins.

Injected as a structured Counselor Preamble directly into the Gemini synthesis prompt. Not a suggestion. A hard constraint.

§ 05 — VISUAL COUNSEL

Walls of dense text repel anxious high schoolers.
We don't output them.

The FastAPI backend returns a Hybrid Structured Response Payload — both structured JSON and minimal, high-nuance markdown blocks. The Next.js frontend reads this payload and renders interactive UI components dynamically.

{ "greeting": "Let's look at Amherst. The money is real here, but the gate is incredibly narrow.", "quick_stats": [ { "label": "Policy Type", "value": "Need-Blind", "badge": "success" }, { "label": "Average Aid", "value": "$66,093", "badge": "info" }, { "label": "Coverage Rate", "value": "77.9%", "badge": "warning" } ], "audit_details": { "the_catch": "Amherst advertises $82k+ average aid. CDS H6 audit shows $66,093 median.", "curriculum_gate": "STATUS: CONDITIONAL. Zero explicit SSC/HSC guidance on record.", "clutch_factor": "Williams guarantees 100% need for all 4 years. Worth more than a larger freshman grant." }, "provenance_hash": "e3bb509dd03e8a84..." }
→ Collapsible accordions for audit details → Spacious metric badges with color coding → Interactive SVG pie charts for aid distribution → Provenance hash in monospace at bottom
§ 06 — VALIDATED OUTPUT

Real response from the deployed backend.
Clinical safety corpus as proof of pattern.

CASE · CLINICAL SAFETY · POST /query
PASS · 0.96 Faithfulness
Query

"What are the rules and guidelines for serious adverse events (SAEs)?"

Greeting

"Let's look at this together. The timeline is strict here."

Quick Stats

Reporting Window: 24 Hours (danger)
Standard AE Logging: 3 Days (warning)

Audit Feedback

Loopholes flagged: expedited reporting loophole closed.
Curriculum warning: none.
Pipeline alert: confirmed.

Provenance hash: e3bb509dd03e8a8498799e80d8aea4a7678b49bfd958d5c9195eb8cc89e2f277

Note: This demo uses the clinical safety corpus as proof of the structured output pattern. Admissions corpus will be populated upon resource activation.

§ 07 — THE GCP STACK

Every Google Cloud service
this project touches.

Discovery Engine (Search API)
Semantic retrieval over raw markdown reports. Credit-safe under App Builder pool.
Cloud Storage
Raw Deep Research Report storage. No database. No schema.
Vertex AI · Gemini Flash
Counselor synthesis with structured JSON output. Temperature 0.0 for deterministic formatting.
Cloud Run
FastAPI backend serving. Autoscaling. Enterprise security headers.
Modal Persistence
Cryptographic provenance chaining. Session state. Append-only NDJSON ledger.
Firebase Auth
JWT verification via ADC. Custom role claims. No JSON key files.
Secret Manager
All credentials. Zero secrets in source.
Cloud Logging
Request traces, audit verdicts, session integrity monitoring.
§ 08 — WHAT THE DOCS DON'T TELL YOU

Five lessons from building Erudite V2.

01
Unstructured beats structured for admissions data

Forcing 70-page forensic reports into JSON schemas destroyed nuance and required endless migrations. Raw markdown + Discovery Engine native parsing preserved everything.

02
Tone is a technical constraint, not a style choice

Injecting the Counselor Preamble as a hard prompt constraint — not a suggestion — transformed output quality. The persona is engineered, not improvised.

03
Visual scannability is the intervention

Anxious 17-year-olds don't read walls of text. Structured JSON → rendered badges/charts/accordions made the same information psychologically accessible.

04
Marketing headlines are dangerous

"$82k average aid" vs "$66k actual median" is a $16k gap that changes enrollment decisions. Forensic CDS auditing is not optional — it's the core value proposition.

05
ADC is non-negotiable

Every Firebase Admin call, every Discovery Engine provisioning request uses Application Default Credentials. Zero JSON key files. Zero credential leaks. Zero deployment friction.

§ 09 — READY ON DAY ONE

Architecture validated.
Corpus loading is the only remaining step.

✓ Done
  • ✅ Backend deployed and health-checked at Cloud Run
  • ✅ Discovery Engine data stores provisioned and tested
  • ✅ Structured output schema validated against live clinical corpus
  • ✅ Modal persistence ledger active with cryptographic chaining
  • ✅ Enterprise security headers hardened (HSTS, XSS, X-Frame)
⏳ Pending resources
  • ⏳ Vanderbilt, Yale, MIT, Amherst, Williams deep research reports
  • ⏳ Bangladesh-specific curriculum contextualization overlays
  • ⏳ Applicant-facing Next.js frontend rendering pipeline
"Every dollar of infrastructure translates directly into structured counsel that prevents a high-need Bangladeshi student from making a $16,000 financial mispricing error. We are ready to begin on day one of resource activation."