← PORTFOLIO RÉSUMÉ
CASE STUDY / 02

ColleAgent

An academic planning prototype that helps students and counselors compare course options. Student records stay private while research agents gather public requirements and source evidence.

TYPEMulti-agent planning
STATUSPre-production alpha
FLOWGoals → Evidence → Plan
FOCUSPrivacy + correctness

01 / PROBLEM

Academic planning needs private context and public evidence at the same time.

A useful planner may need a student's transcript, GPA, schedule, forward plan, major goals, and constraints. But public research agents should not need those sensitive records just to look up articulation rules, program requirements, course equivalencies, deadlines, or college policies.

PRIVATE CONTEXT

Keep student records campus-side

Profile, transcript, GPA, schedule, and forward-plan data remain inside the private planning boundary.

PUBLIC RESEARCH

Send only what research requires

Public agents receive school labels, program/major targets, and bare course codes rather than the student's identity or grades.

PLANNING

Deterministic constraints matter

Requirements, course matching, schedule feasibility, and plan structure should not depend on free-form model improvisation.

HUMAN AUTHORITY

Counselor stays in the loop

The system is designed to support review and correction rather than silently replacing academic advising authority.

02 / PRIVACY ARCHITECTURE

Minimum disclosure is an architectural boundary, not a prompt.

PRIVATE / CAMPUS SIDE

Student planning context

  • identity + profile
  • transcript / grades / GPA
  • schedule + constraints
  • forward plan
  • deterministic planner
MINIMUM DISCLOSURE → fail closed
PUBLIC / RESEARCH SIDE

Evidence gathering

  • college + major
  • public school labels
  • bare course codes
  • catalog / articulation research
  • source evidence + confidence

The Bridge mediates requests between the private planner and public research side. The disclosure packet is intentionally narrow: public research does not require the student's name, GPA, transcript, or detailed grades to answer public academic-policy questions.

03 / SYSTEM FLOW

The user experience is three steps; the system underneath is not.

Goals Capture target colleges, majors, planning horizon, timing, and student constraints on the private side.
Evidence Generate minimum-disclosure research requests, crawl authoritative public sources, and surface exact-major evidence with fail-closed matching.
Plan Combine private student context with researched requirements inside the private planner, then produce ranked schedules and warnings.
Counselor review Keep confidence, evidence gaps, changes, and corrections visible so the plan can be reviewed rather than treated as unquestionable output.

04 / IMPLEMENTATION

Product software around the agent boundary.

CLIENT Next.js Mobile-first Goals → Evidence → Plan flow and signed-session experience.
BRIDGE FastAPI + SQLite Per-route authorization, environment-scoped state, append-only change log, and snapshots.
RESEARCH Public research layer Crawl/sitemap path with Playwright fallback, source evidence, exact-major matching, and failure flags.
PLANNING

Deterministic core

Course matching, requirement handling, feasibility, and ranked plan generation are separated from open-ended research reasoning.

FAILURE HANDLING

Research can fail visibly

Failed or uncertain research is recorded and flagged rather than silently converted into confident planning facts.

AUDITABILITY

Changes are inspectable

Snapshots and append-only changes make planning evolution easier to review and debug.

LONG-TERM DIRECTION

Campus-local agent runtime

The current MVP keeps the campus runtime in-process; a Flower-based campus-local grid is the production direction rather than a current deployment claim.

05 / CURRENT STATE

Alpha software moving toward a counselor-in-loop pilot.

✓
3-step product flowGoals → Evidence → Plan built
✓
Privacy + access controlsminimum-disclosure and authorization paths implemented
✓
Research + planning plumbingcatalog research, fallback browser path, logging, deterministic planning
→
Pilot readinessaccuracy benchmark, computed confidence, staging, load check, sign-in emails

The current goal is not a polished consumer launch. It is a small, observable pilot where research accuracy, privacy boundaries, confidence, and counselor corrections can be measured before broader deployment.

06 / DESIGN PRINCIPLE

Minimal agents. Minimal disclosure. Minimal compute.

ColleAgent deliberately avoids turning every subtask into another autonomous agent. Private planning stays private, public research is scoped to public evidence, and deterministic logic handles the constraints that should be exact.